Apply gini(negatives=...) option; guard degenerate cases

[MaxGhenis]

Summary

Both branches of MicroSeries.gini sorted self directly rather than the local x that was mutated by the negatives option, so negatives='zero' and negatives='shift' were silently ignored. The raw formula with negatives can also return values outside [0, 1], and an all-zero series hit a ZeroDivisionError / RuntimeWarning.

Reproduction

import microdf as mdf

s = mdf.MicroSeries([-5, 0, 10], weights=[1, 1, 1])
s.gini(negatives="zero")    # was 2.0 (unchanged), now 0.6666...
s.gini(negatives="shift")   # was 2.0 (unchanged), now in [0, 1]

mdf.MicroSeries([0, 0, 0], weights=[1, 2, 3]).gini()   # was nan + RuntimeWarning, now 0.0

Fix

• Use the mutated x throughout
• Collapse the two weighted / unweighted branches into a single weighted implementation
• Warn when called with negatives and negatives=None
• Short-circuit all-zero totals to 0 instead of dividing by zero
• Raise on an unknown negatives option

Test plan

• Existing 51 tests still pass
• New test_gini_negatives_option_applied covers negatives=None (warns), 'zero', 'shift', all-zero, and unknown option raising.

[MaxGhenis]

Self-review. Verified:

• x is now used throughout rather than self (the original bug). Both weighted/unweighted branches collapsed into one weighted implementation using np.average-style cumulative trapezoidal area.
• negatives="zero" uses np.where(x<0, 0.0, x) (was buggy in-place which could have ducked the mutated-x bug). negatives="shift" subtracts np.amin(x) only when min < 0. Unknown negatives raises ValueError.
• Degenerate cases: empty → NaN; all-zero total → 0.0 (not NaN/RuntimeWarning); negatives with negatives=None → warning that result may leave [0,1].
• Uses kind="mergesort" for stable sort on ties.
• Test covers warn path, zero clamp (closed-form = 2/3), shift bounded in [0,1], all-zero = 0, invalid arg raises.
• CI green Py 3.9-3.13.