fix: use math.fsum in variance to avoid fp rounding when all credits are equal

[dcondrey]

Summary

• PartialCreditScore.variance was using sum() to compute the mean, which accumulates floating-point rounding error
• When all 8 credits are identical (e.g. 0.8), sum([0.8]*8)/8 is not exactly 0.8, so (c - m)**2 is not exactly 0.0, yielding a residual of ~1.2e-32 instead of 0.0
• Switching to math.fsum (exact summation) for both the mean and the squared-deviations sum eliminates the error

Root cause

0.8 is not exactly representable in IEEE 754; sum() accumulates that error across 8 terms. math.fsum uses an extended-precision accumulator that cancels the error exactly.

Test plan

• pytest tests/test_partial_credit.py passes locally (8/8 green)
• CI passes on all matrix legs (py3.10–3.13 / ubuntu + macos)