Fix stale gradient accumulation in non-AMP training

[sopenlaz0]

Summary

• Clear optimizer gradients after each non-AMP optimizer step.
• Add a regression test exercising two batches through Trainer.train().

Problem

The AMP path calls optimizer.zero_grad(set_to_none=True) after stepping, but
the non-AMP path did not. Consequently, each batch's backward pass accumulated
onto gradients retained from every preceding batch.

With two identical batches and a zero-learning-rate optimizer, the regression
test observes gradients of 1.0 then 2.0 before this fix. After the fix, both
steps observe 1.0, and gradients are cleared after the final step.

Verification

python -m pytest -q tests/test_trainer.py
1 passed