[sharmis-tp]

Summary

Fixes #12 — centered RMSprop can take sqrt of a negative value.

• Clamp the estimated variance to zero before taking the square root, preventing NaN from floating-point cancellation in E[g^2] - E[g]^2 with large gradients (~1e4).
• Move eps outside the sqrt (sqrt(var) + eps instead of sqrt(var + eps)) to match PyTorch and avoid amplifying steps 10000x when variance is near zero.
• Replace the test that asserted runaway step growth with one that verifies step-size convergence and finiteness.
• Add a regression test with 1e4 gradients to guard against the cancellation path.

Test plan

• cargo test — all 42 unit tests, integration tests, and doc-tests pass
• New centered_large_gradients_no_nan test feeds 1e4 gradients and asserts no NaN
• Replaced centered_variant_stabilises_on_constant_gradients verifies convergence