[Gemma4] Fix dtype casting for quantized vision/audio embedders

[sharmax-vikas]

What does this PR do?

Applies the same is_floating_point guard from PR #46904 (which fixed gemma4_unified) to the gemma4 model, which has the same bug.

Bug

In Gemma4VisionPatchEmbedder.forward() and Gemma4AudioSubSampleConvProjectionLayer.forward(), the code does:

pixel_values.to(self.input_proj.weight.dtype)

When loaded with quantization (e.g. BitsAndBytesConfig(load_in_4bit=True)), weight.dtype returns torch.uint8/torch.int8. This silently casts float pixel values to integers, destroying image/audio data. The model produces garbage outputs with no error.

Fix

Add is_floating_point guard (same pattern as #46904):

if (target_dtype := self.input_proj.weight.dtype).is_floating_point:
    pixel_values = pixel_values.to(target_dtype)
hidden_states = self.input_proj(pixel_values)

Fixes #46932

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• - [x] Did you make sure to update the documentation with your changes?
• - [ ] Did you write any new necessary tests?

[github-actions]

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

run-slow: gemma4

[github-actions]

CI Dashboard: View test results in Grafana