[CanReader]

• fix centered RMSprop: clamp variance, move eps outside sqrt (by @sharmis-tp, from Fix centered RMSprop: clamp variance, move eps outside sqrt #73)
• keep eps inside the root and bound the step in the tests

What this changes

Centered RMSprop clamps the variance estimate at zero before the square root, so cancellation in E[g²] - E[g]² can no longer produce a NaN denominator. eps stays inside the root, as before. The two centered tests assert that every step is finite and at or under lr·|g|/√eps.

Why

This is #73 with one correction. The clamp is the fix for #12. Moving eps outside the root, which #73 also did, makes the step lr/eps once the variance estimate reaches zero, which it does exactly in f32 for a constant gradient after about 155 steps: ten thousand times the previous step, the opposite of what that change was meant to do. With eps inside, the step is capped at lr·|g|/√eps, and the tests now check that cap, which the outside placement fails. The contributor's commit is kept with their authorship. Landing from a branch in this repository because the fork's branch cannot be updated from here.

Closes #12.

How it was verified

• cargo test passes
• cargo clippy --all-targets -- -D warnings is clean
• Ran cuda_parity on a GPU
• Trained an example in --release
• No GPU here, CUDA side not runtime tested

The rmsprop unit tests pass locally; CI runs the rest.

Checklist

• Every new or changed backward rule has a gradcheck case
• Every new kernel has a cuda_parity case
• tensor/ops and autograd/ops still mirror each other
• Public items have /// docs