test(pp): fail parity tests when neither run trained (#3701)

The parity check only compares the two runs against each other, so two equally
broken runs agree perfectly and pass. A qwen3_5_moe proxy hit exactly that: its
weights were silently left uninitialized, both legs reported zero gradients and
a loss stuck at ln(vocab_size), and the comparison called it a pass.

Fail instead when a run reports grad_norm of exactly 0, or when the baseline
loss never moves across steps that each saw a different batch.

Signed-off-by: Abhishree <abhishreetm@gmail.com>
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>