fix(r_squared): implement the documented zero_division parameter

[hassaanch23]

Closes #803.

The problem

r_squared documents a zero_division parameter that was never implemented:

zero_division: Which value to substitute as a metric value when encountering zero division.
    Should be one of 0, 1, "warn". "warn" acts as 0, but the warning is raised.

zero_division appeared exactly once in the file — in that docstring. _compute didn't accept it, so passing it raised TypeError, and the case it exists to control was unguarded:

sst = np.sum((references - mean_references) ** 2)
r_squared = 1 - (ssr / sst)

When every reference is identical there is no variance to explain, sst is zero, and R^2 is undefined. numpy returns -inf — or nan when predictions match exactly — with only a RuntimeWarning, and round() passes it through.

predictions	references	before	after (default)	sklearn.r2_score
[1.1, 2.1, 2.9, 4.2]	[1, 2, 3, 4]	0.986	0.986	0.986
[1, 2, 3, 4]	[5, 5, 5, 5]	-inf	0.0 + warning	0.0
[5, 5, 5, 5]	[5, 5, 5, 5]	nan	0.0 + warning	1.0

Both values propagate: one constant-reference batch turns an averaged score into nan for a whole run, and -inf dominates any mean. Nothing raises, so it reads as a modelling outcome rather than an undefined computation.

The change

Implements zero_division as documented — 0, 1, or "warn" — and raises ValueError on anything else rather than ignoring it.

The default is "warn": the previously silent case becomes visible instead of quietly becoming a number the caller didn't choose. Callers who know which convention they want can pass 0 or 1 and silence it. Note the two degenerate cases differ under sklearn (0.0 when predictions miss, 1.0 when they match exactly), which is exactly why this is a caller choice rather than a fixed substitution.

Non-degenerate inputs are untouched and still agree with sklearn.metrics.r2_score to 3 dp.

Tests

Adds doctest examples for both settings, which is how metrics here are covered. tests/test_metric_common.py -k r_squared passes (1 passed, 1 skipped).

One note for anyone reproducing: the test suite doesn't run on Python 3.12+ — tests/utils.py imports distutils, removed in 3.12. Verified on 3.11.

Not changed

_DESCRIPTION says "The R^2 value ranges from 0 to 1", which isn't true for a fit worse than the mean — sklearn returns negative values there, as does this implementation. Left alone to keep this PR to one thing; happy to fix it here or separately.