fix: prevent LM damping underflow

[jlperla]

Summary

• add min_damping as a traced LM hyperparameter
• default the floor to jnp.finfo(residual.dtype).tiny
• clamp damping before every linear solve and after each accept/reject update
• document the distinction between an underflow floor and an eps * operator_scale conditioning floor
• cover eager/JIT float32, LSMR, explicit floors, static cache keys, and float64 behavior

Root cause

Repeated accepted steps could push damping into a backend's flush-to-zero range. Once damping became exactly zero, the multiplicative update could not recover it. Rank-deficient factorizations could then produce rejected NaN trials indefinitely, while LSMR divided by sqrt(damping) directly.

Impact

The default now prevents literal damping underflow without imposing a materially larger regularization. Users can set a larger absolute min_damping when they need a conditioning-oriented floor.

Validation

• uv run pytest -q — 565 passed, 12 skipped
• uv run ruff check .
• uv run --group docs mkdocs build --strict