fix(testing): Fix MoonshineEncoder UnboundLocalError and Florence2VisionBackbone dtype mismatch

[harshaljanjani]

What does this PR do?

The following failing tests were identified and fixed in this PR:

→ Moonshine: In MoonshineEncoder.forward, the var output_attention_mask is only assigned inside the if attention_mask is not None clause, but create_bidirectional_mask called immediately after with attention_mask=None still returns a non-None tensor. Under torch.compile(dynamic=True), is_tracing() returns True inside _ignore_bidirectional_mask_sdpa → not is_tracing(padding_mask) fails and _ignore_bidirectional_mask_sdpa returns False, so the allow_is_bidirectional_skip early-exit in sdpa_mask clause never fires and a full batched tensor is materialized (shown in the trace). Speaking to the return-time guard, attention_mask is not None becomes truthy even when no padding mask was provided, leaving output_attention_mask unset. The aforementioned case is hit by test_sdpa_can_compile_dynamic since it drops all attention masks before calling the model and causes an UnboundLocalError; this change should fix that.

Here are the trace outputs that led to arrive at the above explanation:

→ MoonshineEncoder.forward:   input attention_mask=None
→ MoonshineEncoder.forward:   attention_mask is None; output_attention_mask NOT assigned
→ MoonshineEncoder.forward:   calling create_bidirectional_mask with attention_mask=None
→ _preprocess_mask_arguments: early_exit=False, returning attention_mask=None, q_length=1, kv_length=1
→ create_bidirectional_mask:  mask_interface=sdpa_mask
→ create_bidirectional_mask:  calling mask_interface with allow_is_bidirectional_skip=True, attention_mask=None
→ create_bidirectional_mask:  mask_interface returned type=Tensor, shape=torch.Size([3, 1, 1, 1])
→ MoonshineEncoder.forward:   guard 'attention_mask is not None' = True
→ MoonshineEncoder.forward:   about to return, evaluating guard now
>> **UnboundLocalError**

→ Florence2Vision: test_sdpa_can_compile_dynamic loads the model in bfloat16, but Florence2VisionBackbone.forward got float32 activations from upstream processing but (its conv weights self.convs[0].conv.weight) were bfloat16. Casting the dtype fixes this :)

CI Failures:

Before the fix (feel free to cross-check; these errors are reproducible):

After the fix (feel free to cross-check):

cc: @Rocketknight1

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you fix any necessary existing tests?

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: florence2, moonshine

[Rocketknight1]

Yes, LGTM!

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.