[Gemma3n] Fix FA2 vs eager equivalence tests
Gemma3n's last hidden states are large enough that bf16 rounding alone exceeds the common 3e-2 tolerance (eager
vs sdpa drift the same way, fp32 matches to ~1e-6), and its `hidden_states` have a leading altup dim that the
common padded check slices as the batch, comparing the padded token. Compare `last_hidden_state` on non-padded
tokens with bf16-appropriate tolerances instead.
Fixes #50015
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>