feat: Repr prints exactly what NumPy 2.5.3's repr prints

[tannevaled]

Repr wrote each element in its shortest form and right-aligned them, so its digits and padding differed from NumPy's:

array	before	NumPy 2.5.3 (now)
[1.5, 2, 3] float32	array([1.5, 2., 3.], dtype=float32)	array([1.5, 2. , 3. ], dtype=float32)
[1e20, 0.5]	array([1e+20, 0.5])	array([1.e+20, 5.e-01])
[True, True]	array([True, True])	array([ True, True])
arange(2000.)	array([ 0., 1., …	array([0.000e+00, 1.000e+00, … (max/min > 1000)

display.go is now a port of numpy/_core/arrayprint.py at v2.5.3:

• Formatters: FloatingFormat (maxprec, precision 8), ComplexFloatingFormat, IntegerFormat and BoolFormat.
• Layout: _formatArray and _extendLine with 75 columns.
• Suffixes: the shape= and dtype= extras, including their wrapping onto a new line.
• Digits: taken from dragon4.c: the shortest that read back, cut to 8 and rounded half to even on the exact value. That is also how strconv rounds.
• Scientific cut-off: it depends on the dtype, as in 2.5.3: 1e8 for float64, 1e6 for float32.

Oracle. NumPy's own expected strings, copied verbatim from test_arrayprint.py (v2.5.3):

• 26 repr cases, from test_nan_inf, test_float_spacing, test_bool_spacing, test_sign_spacing, test_floatmode (maxprec), test_summarize_1d/2d and test_summarize_2d_dtype;
• 72 complex special values (TestComplexArray.test_str, complex64 and complex128);
• gh-28679's float32/complex64 cut-off elements.

Four more cases have no default-options case in NumPy's suite (empty arrays, the 8-digit cut in scientific mode, a negative tiny value). For those, the expectation is derived from the rules, and the rule is written beside each case.

Checks:

• Setting the float32 cut-off back to 1e8 fails the tests, and so does dropping the bool padding.
• Three old expectations encoded the previous, wrong output; they are corrected, with the reason in a comment.
• Coverage is 100% for the root package.
• All packages pass, and the tests compile for linux/386 and arm.

Downstream: go-embedded-ruby's NDArray#inspect prints this Repr. Its tests that compare against go-ndarray computed in Go will follow on their own. Literal strings in its tests will need the NumPy forms when it moves to this release.

🤖 Generated with Claude Code