Fix Pfaffian backward for numerically singular inputs

[JeremieGince]

Summary

• Replace the exact pf == 0 singularity check in PfaffianStrategy's backward pass with two dtype-adaptive criteria: a relative-magnitude test on the Pfaffian and a residual check on the LU inverse, so numerically (not just exactly) singular or ill-conditioned matrices are routed to the exact minor-based adjugate instead of producing garbage or NaNs from the LU inverse.
• Add SINGULARITY_RTOL_EXPONENT and INVERSE_RESIDUAL_RTOL_EXPONENT class constants that scale with torch.finfo(dtype).eps, keeping the thresholds correct across float32/float64.
• Add test coverage for moderate ill-conditioning, ties between the two criteria, mixed-batch routing, and dtype adaptivity.

Test plan

• pytest tests/test_strategies/test_strategy.py — 14 passed, 100% coverage on strategy.py
• Full suite: pytest — 495 passed, 100% coverage

🤖 Generated with Claude Code

[github-actions]

☂️ Python Coverage

current status: ✅

Overall Coverage

Lines	Covered	Coverage	Threshold	Status
280	280	100%	98%	🟢

New Files

No new covered files...

Modified Files

File	Coverage	Status
src/torch_pfaffian/strategies/strategy.py	100%	🟢
TOTAL	100%	🟢

updated for commit: ae1af29 by action🐍