Skip zero-weight rows in MicroSeries.quantile

[MaxGhenis]

Summary

MicroSeries.quantile could select a zero-weight element because np.searchsorted(cumsum_norm, 0, side='left') always returns 0, and the same class of bug fired at internal plateaus of zero weight. The fix drops zero-weight rows before sorting and computing the CDF; an all-zero-weight series returns NaN.

Reproduction

import microdf as mdf

MicroSeries([10, 20, 30], weights=[0, 1, 1]).quantile(0.0)  # was 10, now 20

Test plan

• Existing 51 tests still pass
• New test_quantile_skips_zero_weight_rows covers:
  • Zero-weight first row
  • Internal zero-weight plateau (the zero-weight value must never be selected across q in [0, 1])
  • All zero weights → NaN

[MaxGhenis]

Self-review. Verified:

• nonzero = sample_weight > 0 filter applied before the CDF search; if no rows have positive weight, return np.nan (scalar q) or a NaN-filled Series (array q) — documented behaviour.
• Fix addresses both the q=0 case and internal zero-weight plateaus because filtering happens before argsort/cumsum.
• Regression test covers q=0 with first-row zero weight, internal plateau (sweeping np.linspace(0,1,21) to assert the zero-weight value never appears), and all-zero-weight → NaN.
• Does not perturb the correct-weight path (all tests still pass).
• CI green Py 3.9-3.13.