Fix dtype cast in TimesFM 2.5 ResidualBlock for quantized weights

[JOhnsonKC201]

What does this PR do?

TimesFm2_5ResidualBlock.forward unconditionally casts its float input to the block's weight dtype:

x = x.to(self.input_layer.weight.dtype)

input_layer is an nn.Linear, and this block is used as the model's input projection (input_ff_layer), so the raw float time-series inputs flow through it. When the model is loaded with quantization (e.g. bitsandbytes 4/8-bit), input_layer.weight.dtype is an integer type (e.g. uint8). Casting the float inputs to an int dtype silently truncates them to zeros (0.73 -> 0), so the model produces garbage outputs with no error or warning.

The fix only casts when the weight dtype is floating point, preserving the original mixed-precision behavior while skipping the destructive cast under quantization:

if (target_dtype := self.input_layer.weight.dtype).is_floating_point:
    x = x.to(target_dtype)

This is the same pattern already applied to the multimodal embedders in gemma4_unified (#46904) and gemma4 (#46933).

Both modular_timesfm2_5.py and the generated modeling_timesfm2_5.py are updated to stay in sync.

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?
• Was this discussed/approved via a GitHub issue? Please add a link to it if that's the case. (Same class of bug as Fix dtype casting for quantized multimodal embedders #46904 / [Gemma4] Fix dtype casting for quantized vision/audio embedders #46933.)
• Did you make sure to update the documentation with your changes? N/A.
• Did you write any new necessary tests? Follows the precedent of Fix dtype casting for quantized multimodal embedders #46904/[Gemma4] Fix dtype casting for quantized vision/audio embedders #46933 (dtype-safety guard, no new test).

Who can review?

@kashif @SunMarc

[github-actions]

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

run-slow: timesfm2_5

[github-actions]

CI Dashboard: View test results in Grafana

[JOhnsonKC201]

Small, self-contained fix (+10/-4, with modular_timesfm2_5.py and the generated modeling_timesfm2_5.py kept in sync). The destructive cast only triggers under quantization, where input_layer.weight.dtype is an integer type — so the float time-series inputs get silently truncated to zeros with no error in normal fp runs. Same failure mode and fix as the multimodal embedders in #46904 and #46933.

Happy to run run-slow: timesfm2_5 or adjust anything once a maintainer has a chance to look. Thanks!

[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.