Refactor GGUF to speed-up inference

[SunMarc]

What does this PR do?

This PR updates the GGUF in transformers in order to take advantages of some features that landed for the past few months and are now stable:

• conversion using WeightConverter, instead of our custom tensor processor.
• Support for loading quantized models. Before we were dequantizing by default.
• Add support llamacpp kernels to speed up inference for quantized models (e.g q4, q6) on CUDA and MacOS with kernels

For now, I'm restricting myself to add support only qwen3.5 with this new API but I will add the rest once this lands and remove the old API. Right now, it will fallback to the old one if the arch is not supported.

Todo:

• Publish CUDA/Metal kernels and add kernels.py file
• Add some benchmarks
• Common tests for GGUF integration

[LysandreJik]

Thanks! cc @ArthurZucker fyi (still actively wip, no need to review 🤗)

[github-actions]

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

run-slow: ggml

[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.

[github-actions]

---

CI recap

Dashboard: View test results in Grafana
Latest run: 30941170812:1
Result: failure | Jobs: 2 | Tests: 7 | Failures: 1 | Duration: 17s