Fix inactive JAX diagonal cells poisoning path gradients

[geekbeast]

Summary

Prevent inactive lanes in the fixed-width JAX diagonal sweep from overflowing and poisoning native JAX gradients. The final kernel value can be finite while unused internal states produce NaNs during backpropagation; outer gradient clipping cannot fix that.

• Mask boundary vectors and path operands before static-kernel/power-series arithmetic; use safe gather indices and where for boundary selection.
• Apply the same rule to chunked and unchunked forward sweeps.
• Use row-consistent rectangular geometry so the active-lane mask is correct for unequal path lengths.
• Remove mutation of self.exponents inside the jitted single-pair entry point, which leaked a tracer on reuse in the new differentiation tests.
• Add synthetic FP32 regressions and an independent row-major reference; wire them into CPU CI.

No custom-VJP implementation is imported from another repository. No dataset samples, learned weights, preprocessing changes, gradient clipping changes, or dtype changes are included. Real active-tile overflow remains visible.

Validation

• Before fix: a dataset-free 64-knot linear example at order 9 returns a finite kernel value but 46 non-finite query-gradient entries and 2 non-finite bank-gradient entries.
• CPU: pytest tests/test_core_jax.py tests/test_diagonal_masking.py -q — 46 passed.
• GPU: pytest tests/test_diagonal_masking.py -q — 25 passed; strict FP32, highest matmul precision, NVIDIA_TF32_OVERRIDE=0.
• Tests cover linear/RBF kernels, both rectangular orientations, one-increment paths, chunking, poisoned inactive inputs, independent gradient agreement and active overflow preservation.

See docs/diagonal-masking.md for the numerical rationale and stress-test tolerance.