Speed up image comparison

[akx]

Closes #7260 (supersedes it).

I added a benchmark, and stacked the tests from the original (adjusted to pass lints, etc.) to be before the reimplementation in the PR, then added some more corner case tests.

I verified, with interactive rebase, that the tests pass just before the new handler is added, and then separately afterwards.

Local benchmark score improvement (unsurprising, and this isn't measuring peak memory usage improvements!):

------------------- benchmark 'compare': 15 tests, 2 sources -------------------
Name (time in ns)                          0001_main Min  0002_new Min      ΔMin
--------------------------------------------------------------------------------
test_equality[1237x811-L-inverted]          122,875.0916      470.7952    -99.6%
test_equality[1237x811-L-equal]             125,959.0499   24,207.9841    -80.8%
test_equality[1237x811-L-one-pixel]         129,332.9988   12,415.9269    -90.4%
test_equality[1237x811-I;16-one-pixel]      343,374.9080   27,542.0025    -92.0%
test_equality[1237x811-LA-inverted]         343,417.0503      479.1538    -99.9%
test_equality[1237x811-I;16-equal]          380,958.0812   55,041.9791    -85.6%
test_equality[1237x811-I;16-inverted]       387,958.9494      479.0956    -99.9%
test_equality[1237x811-LA-one-pixel]        451,375.0318  189,000.0422    -58.1%
test_equality[1237x811-LA-equal]            463,792.0065  378,541.9976    -18.4%
test_equality[1237x811-RGBA-inverted]       727,667.0076      483.2982    -99.9%
test_equality[1237x811-RGB-inverted]        754,792.0104      479.2004    -99.9%
test_equality[1237x811-RGB-equal]           813,374.9943  378,374.9416    -53.5%
test_equality[1237x811-RGB-one-pixel]       823,125.0104  189,290.9640    -77.0%
test_equality[1237x811-RGBA-equal]          874,208.0536  105,833.0527    -87.9%
test_equality[1237x811-RGBA-one-pixel]      899,208.0111   52,874.9079    -94.1%
--------------------------------------------------------------------------------

EDIT: added one more commit for autovectorisation to make RGB (likely very common!) and LA even faster (at least on my hardware):

--------------------- benchmark 'compare': 15 tests, 2 sources ---------------------
Name (time in ns)                          0001_main Min  0005_equalve Min      ΔMin
------------------------------------------------------------------------------------
test_equality[1237x811-L-inverted]          122,875.0916          479.0956    -99.6%
test_equality[1237x811-L-equal]             125,959.0499       24,333.0142    -80.7%
test_equality[1237x811-L-one-pixel]         129,332.9988       12,416.9746    -90.4%
test_equality[1237x811-I;16-one-pixel]      343,374.9080       27,625.0066    -92.0%
test_equality[1237x811-LA-inverted]         343,417.0503          479.2004    -99.9%
test_equality[1237x811-I;16-equal]          380,958.0812       55,207.9873    -85.5%
test_equality[1237x811-I;16-inverted]       387,958.9494          479.2004    -99.9%
test_equality[1237x811-LA-one-pixel]        451,375.0318       43,249.9219    -90.4%
test_equality[1237x811-LA-equal]            463,792.0065       86,874.9339    -81.3%
test_equality[1237x811-RGBA-inverted]       727,667.0076          485.3995    -99.9%
test_equality[1237x811-RGB-inverted]        754,792.0104          483.2982    -99.9%
test_equality[1237x811-RGB-equal]           813,374.9943       86,790.9985    -89.3%
test_equality[1237x811-RGB-one-pixel]       823,125.0104       43,459.0038    -94.7%
test_equality[1237x811-RGBA-equal]          874,208.0536      105,957.9663    -87.9%
test_equality[1237x811-RGBA-one-pixel]      899,208.0111       52,916.9338    -94.1%
------------------------------------------------------------------------------------

[akx]

CodSpeed isn't showing a report here because all the benchmarks are new, but you can see them in the report here.

The varied range of speeds suggests that these optimizations do take, but I'm not sure if the last commit, the autovectorisation, works as well as it does on clang/ARM.

[hugovk]

Want to split the benchmarks to their own PR?

[akx]

@hugovk Done: #10033

[codspeed]

Merging this PR will improve performance by ×32

⚠️ Different runtime environments detected

Some benchmarks with significant performance changes were compared across different runtime environments,
which may affect the accuracy of the results.

Open the report in CodSpeed to investigate

⚡ 15 improved benchmarks
✅ 609 untouched benchmarks
⏩ 338 skipped benchmarks¹

Performance Changes

	Benchmark	BASE	HEAD	Efficiency
⚡	test_equality[1237x811-RGB-inverted]	49,211 µs	67.4 µs	×730
⚡	test_equality[1237x811-RGBA-inverted]	52,269.3 µs	73.7 µs	×710
⚡	test_equality[1237x811-LA-inverted]	30,972.3 µs	67.6 µs	×460
⚡	test_equality[1237x811-I;16-inverted]	25,080.7 µs	67.5 µs	×370
⚡	test_equality[1237x811-L-inverted]	12,659.3 µs	67.1 µs	×190
⚡	test_equality[1237x811-RGBA-one-pixel]	54.6 ms	3.7 ms	×15
⚡	test_equality[1237x811-I;16-one-pixel]	25.7 ms	1.9 ms	×13
⚡	test_equality[1237x811-L-one-pixel]	12,877.2 µs	985.3 µs	×13
⚡	test_equality[1237x811-RGB-one-pixel]	51.3 ms	4 ms	×13
⚡	test_equality[1237x811-LA-one-pixel]	32.1 ms	4 ms	×8.1
⚡	test_equality[1237x811-RGBA-equal]	57.7 ms	7.4 ms	×7.8
⚡	test_equality[1237x811-I;16-equal]	26.3 ms	3.7 ms	×7
⚡	test_equality[1237x811-L-equal]	13.1 ms	1.9 ms	×6.9
⚡	test_equality[1237x811-RGB-equal]	52.7 ms	7.9 ms	×6.7
⚡	test_equality[1237x811-LA-equal]	33.2 ms	7.9 ms	×4.2

Tip

Curious why performance improved? Comment @codspeedbot explain why performance improved on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing akx:image-equals-new (8602567) with main (d13f288)}

Footnotes

1. 338 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

[akx]

And the memory usage comparison that the benchmark doesn't capture shows that comparisons now require no extra allocations:

On a 2053x2039 image:

Mode	main == (MiB)	new == (MiB)	main peak (MiB)	new peak (MiB)
1	3.7	0.0	11.8	8.1
CMYK	57.7	0.0	89.8	32.0
F	56.8	0.0	88.8	32.0
HSV	42.0	0.0	74.0	32.0
I	56.8	0.0	88.8	32.0
I;16	28.7	0.0	44.8	16.0
I;16B	39.9	0.0	55.9	16.0
I;16L	29.5	0.0	45.5	16.0
I;16N	28.7	0.0	44.8	16.0
L	15.2	0.0	23.3	8.1
LA	28.7	0.0	60.7	32.0
LAB	53.8	0.0	85.9	32.1
La	28.7	0.0	60.7	32.0
P	16.9	0.0	25.2	8.2
PA	28.8	0.0	61.0	32.1
RGB	42.0	0.0	74.0	32.0
RGBA	56.8	0.0	88.8	32.0
RGBX	56.8	0.0	88.8	32.0
RGBa	56.8	0.0	88.8	32.0
YCbCr	42.0	0.0	74.0	32.0