numba: inline="always" to speedup trivial op compilation

[ricardoV94]

This speedups first time compilation. Numba prefers to compile a single function with may operations instead of our - one function per op - design. Added inline=always to trivial cheap Ops (scalar, dimshuffle, reshape, ...)

Note the speedup in compile with cache=False (False flag). With cache=True we are seeing the dominating effect of compiling something we already saw before, that isn't changed. But the first time compile should now be faster as well.

Before
------------------------------------------------------------------------------------------------------------------------------- benchmark: 13 tests -------------------------------------------------------------------------------------------------------------------------------
Name (time in us)                                                          Min                        Max                       Mean                    StdDev                     Median                       IQR            Outliers           OPS            Rounds  Iterations
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
test_radon_model_call_benchmark[NUMBA-False]                            7.1030 (1.0)             508.6630 (14.56)             7.4571 (1.0)              2.1671 (inf)               7.3640 (1.0)              0.1310 (inf)      573;3968  134,100.2381 (1.0)       57003           1
test_radon_model_call_benchmark[NUMBA-True]                             7.2030 (1.01)             34.9450 (1.0)               7.5947 (1.02)             0.9135 (inf)               7.4540 (1.01)             0.0900 (inf)     1984;2964  131,670.2561 (0.98)      55605           1
test_radon_model_compile_repeatedly_benchmark[NUMBA-True]         254,713.1150 (>1000.0)   5,534,446.4740 (>1000.0)   1,396,182.6770 (>1000.0)  2,320,576.2290 (inf)         257,882.6410 (>1000.0)  1,637,049.1070 (inf)           1;1        0.7162 (0.00)          5           1
test_radon_model_compile_repeatedly_benchmark[NUMBA-False]      5,671,687.5410 (>1000.0)   8,076,519.9060 (>1000.0)   6,710,889.8018 (>1000.0)  1,012,906.0587 (inf)       6,619,624.3880 (>1000.0)  1,725,231.5448 (inf)           2;0        0.1490 (0.00)          5           1
test_radon_model_compile_variants_benchmark[NUMBA-True]        20,028,947.0980 (>1000.0)  20,028,947.0980 (>1000.0)  20,028,947.0980 (>1000.0)          0.0000 (1.0)      20,028,947.0980 (>1000.0)          0.0000 (1.0)           0;0        0.0499 (0.00)          1           1
test_radon_model_compile_variants_benchmark[NUMBA-False]       50,546,677.2310 (>1000.0)  50,546,677.2310 (>1000.0)  50,546,677.2310 (>1000.0)          0.0000 (1.0)      50,546,677.2310 (>1000.0)          0.0000 (1.0)           0;0        0.0198 (0.00)          1           1
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------

After
------------------------------------------------------------------------------------------------------------------------------- benchmark: 13 tests -------------------------------------------------------------------------------------------------------------------------------
Name (time in us)                                                          Min                        Max                       Mean                    StdDev                     Median                       IQR            Outliers           OPS            Rounds  Iterations
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
test_radon_model_call_benchmark[NUMBA-False]                            6.6530 (1.0)              99.5560 (1.15)              7.3356 (1.0)              1.4372 (inf)               7.2330 (1.0)              0.1500 (inf)      862;3896  136,322.0726 (1.0)       49856           1
test_radon_model_call_benchmark[NUMBA-True]                             7.2030 (1.08)          1,034.4090 (11.91)             7.5908 (1.03)             4.4430 (inf)               7.4740 (1.03)             0.1000 (inf)      286;2677  131,738.6749 (0.97)      54813           1
test_radon_model_compile_repeatedly_benchmark[NUMBA-True]         258,900.6500 (>1000.0)   5,994,141.2680 (>1000.0)   1,408,531.9982 (>1000.0)  2,563,436.1926 (inf)         261,233.4620 (>1000.0)  1,440,107.1928 (inf)           1;1        0.7100 (0.00)          5           1
test_radon_model_compile_repeatedly_benchmark[NUMBA-False]      5,084,492.1370 (>1000.0)   7,516,582.0820 (>1000.0)   5,725,177.9268 (>1000.0)  1,010,872.8182 (inf)       5,326,797.6500 (>1000.0)    779,492.7115 (inf)           1;1        0.1747 (0.00)          5           1
test_radon_model_compile_variants_benchmark[NUMBA-True]        20,643,653.0320 (>1000.0)  20,643,653.0320 (>1000.0)  20,643,653.0320 (>1000.0)          0.0000 (1.0)      20,643,653.0320 (>1000.0)          0.0000 (1.0)           0;0        0.0484 (0.00)          1           1
test_radon_model_compile_variants_benchmark[NUMBA-False]       42,041,049.0130 (>1000.0)  42,041,049.0130 (>1000.0)  42,041,049.0130 (>1000.0)          0.0000 (1.0)      42,041,049.0130 (>1000.0)          0.0000 (1.0)           0;0        0.0238 (0.00)          1           1
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------

[jessegrabowski]

nice and straight-forward. Win seems marginal but we take those.

[ricardoV94]

gotta wait, we had a big regression in scan... without #2098