[FEATURE SUPPORT] Add HuggingFace Kernel Hub support

[LoserCheems]

Summary

• Add support for publishing Flash Sparse Attention Triton kernels to HuggingFace Kernel Hub, enabling users to load kernels directly via
  get_kernel() without installing the full package.

Design

• A build script (scripts/build_hf_kernels.py) generates a self-contained kernel package under huggingface_kernels/, rewriting absolute
  imports to relative imports and producing all required config files (build.toml, flake.nix, CARD.md).
• The package is built and uploaded using kernel-builder from HuggingFace, which uses Nix for reproducible sandboxed builds.
• Alternative considered: manual packaging without kernel-builder. Rejected because kernel-builder handles ABI compatibility across
  CUDA/PyTorch versions automatically.

Changes

• scripts/build_hf_kernels.py — New script to generate the HuggingFace kernel package with import rewriting, config generation, and compliance
  checks.
• README.md, README_zh.md — Added HuggingFace Kernel usage section under Installation.
• docs/huggingface_kernels_guide.md — Full build & publish guide.
• mkdocs.yml — Added guide to docs navigation.

Implementation Notes

• The Nix binary cache address is huggingface.cachix.org (not huggingface-kernels.cachix.org).
• kernel-builder build-and-upload may fail with 403 if the repo already exists. The recommended workflow is build-and-copy followed by upload
  --repo-type model.
• NIX_CONFIG can only be exported once; multiple exports overwrite each other.

Tests

• scripts/test_hf_kernel.py validates loading from Hub and runs dense forward, sparse forward, and dense decode.
• huggingface_kernels/tests/test_flash_attn.py is generated by the build script for kernel-builder's CI.
• All 12 public functions verified present in the generated init.py.

Docs

• docs/huggingface_kernels_guide.md — Build & publish guide.
• README.md / README_zh.md — Usage examples added.
• mkdocs.yml — Navigation updated.