Optimize Parakeet feature extraction on CUDA

[milesial]

What does this PR do?

Add support for CUDA parakeet preprocessor, running STFT and mel spectrogram extraction on the GPU.
This refactor also speeds up the CPU implementation.

Tested on nvidia/parakeet-ctc-0.6b, B200, 300s audio:

Before this PR, CPU: 28ms
After this PR, CPU: 21ms
After this PR, GPU: 1.7ms

No impact on accuracy (VoxPopuli).

Context for this one is to accelerate vLLM for our multimodal nemotron model. Processing audio inputs has a bottleneck on this CPU feature extractor. This PR does a small refactor that gives some good speedup for CPU, and also enables the CUDA backend to be used to accelerate even further.

Some rough numbers, processing a 30min audio of 16 kHz:

Before this PR, CPU: 770ms
After this PR, CPU: 265ms (~3x)
After this PR, CUDA: 20ms (~38x)

Several changes in this PR:

Enabling CUDA backend
Caching of the Hann window and mel filters
Dynamic torch compile at strategic locations
GPU-friendly padding
More efficient and accurate complex number magnitude computation

• I confirm that this is not a pure code agent PR.

Before submitting

• Did you read the contributor guideline,
  Pull Request section?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

[Rocketknight1]

cc @eustlb @ebezzam

[eustlb]

Hey @milesial, interesting PR! This should be covered out of the box by #44394, so I'll wait it to land (coming days) beofre doing a full review. Context on what motivated you to do so and limitations you had with the current implem would be golden feedback for us 🙏

[milesial]

Hi @eustlb , thanks for linking your PR!
Context for this one is to accelerate vLLM for our multimodal nemotron model. Processing audio inputs has a bottleneck on this CPU feature extractor. This PR does a small refactor that gives some good speedup for CPU, and also enables the CUDA backend to be used to accelerate even further.

Some rough numbers, processing a 30min audio of 16 kHz:

• Before this PR, CPU: 770ms
• After this PR, CPU: 265ms (~3x)
• After this PR, CUDA: 5.5ms (~140x)

Several changes in this PR:

• Enabling CUDA backend
• Caching of the Hann window and mel filters
• Dynamic torch compile at strategic locations
• GPU-friendly padding
• More efficient and accurate complex number magnitude computation

I see your PR is a major refactor of audio processors, does it address these points as well?
I tried your PR and these are the numbers I got:

• CPU: 690ms
• CUDA (after some patches, did not work OOTB): 16.5ms

Other question would be about timelines, I'm guessing your PR can take a while to get merged, while this smaller one can get merged faster in the meantime and unblock us, what do you think?

[github-actions]

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

run-slow: parakeet

[milesial]

Gentle ping @eustlb @ebezzam on the above comment