feat(rocm): enable native GGUF kernels on RDNA4

[zihaomu]

Draft follow-up to #132 and #133. This branch is based on the JIT follow-up head. Until its dependencies merge, GitHub Files changed also includes their diffs; the incremental GGUF review scope is the five files listed below.

Summary

• stage GGUF CUDA sources in the extension cache before HIPify, keeping the checkout unchanged;
• key the staged source cache by source content and Torch/HIP versions;
• add ROCm runtime link flags and select the backend-neutral Thrust path required by the generic Ubuntu headers;
• use the 64-bit synchronized-shuffle mask required by HIP, including wave32 RDNA4 targets;
• document the optional Thrust-header dependency and add a Q4_0 dequantization test against the pure-Torch reference.

Incremental review scope

Relative to zihaomu:feat/rocm-rdna4-jit:

• python/freetoken/kernel/gguf.py
• python/freetoken/kernel/csrc/gguf/dispatch.h
• tests/kernels/test_gguf_rocm.py
• docs/install.md
• README.md

Diff size: 97 additions, 8 deletions.

Dependency

• Depends on feat(rocm): add RDNA3 and RDNA4 runtime foundation #132 for the ROCm/RDNA4 foundation.
• Depends on fix(rocm): make TVM-FFI index and store JIT kernels portable to HIP #133 for common HIP JIT launch/link support.
• Does not depend on the backend or RCCL follow-ups.

Validation

Hardware and software:

• AMD Radeon AI PRO R9700 (gfx1201)
• PyTorch 2.11.0+rocm7.14.0, HIP 7.14.60850

Cold extension-cache result:

• full source HIPify: 187 kernel launches replaced;
• unsupported CUDA function calls: 0;
• hipcc compilation and shared-library link: PASS;
• Q4_0 native dequantization vs pure-Torch reference: PASS;
• cold build/test duration: 124 seconds.

Known limitations

Only the Q4_0 dequantization path is claimed numerically here. Other GGUF quant formats and full model serving remain out of scope. This PR remains Draft until #132 and #133 establish the final base.