feat: add simple ROCm GPU support for RDNA3 (gfx1100-1103)

[bouclem]

ROCm support for RDNA3 GPUs

Adds initial ROCm support so FreeToken builds and runs on AMD RDNA3 cards (gfx1100-1103).

Supported GPUs:

• gfx1100 - RX 7900 XTX / XT
• gfx1101 - RX 7900 GRE
• gfx1102 - RX 7700 / XT
• gfx1103 - RX 7600 / XT

What works:

• pinned_tensor.cpp compiles under hipcc via a hip_compat.h shim that maps CUDA runtime API to HIP
• setup.py detects ROCm at build time, links amdhip64, passes --offload-arch for the four gfx targets
• JIT compilation routes to HIP flags on ROCm
• NVIDIA-specific arch checks (sm90/sm100) return None on ROCm so the engine falls back to Triton kernels
• nvcc version check is skipped on ROCm
• clangd generation falls back to rocm-smi

Left as TODO(ROCm) in the code:

• NCCL to RCCL migration for multi-GPU
• flashinfer/sgl_kernel have no ROCm builds (Triton fallback works)
• Triton autotune configs need RDNA3 tuning
• PDL has no ROCm equivalent
• rocm-smi auto gfx detection (currently relies on env vars)

[hakatu]

Could add ROCm support for RDNA4 GPUs as well?

[zihaomu]

For RDNA4 users, the complete validated adaptation is available on zihaomu:feat/rdna4-rocm. The branch has been restored and currently points to commit 3b51956.

This is the monolithic, runnable RDNA4 snapshot used for validation on an AMD Radeon AI PRO R9700 (gfx1201) with PyTorch 2.11.0+rocm7.14.0, HIP 7.14.60850, and AMD Triton 3.7.1. It includes the ROCm build/runtime foundation, TVM-FFI JIT portability, ROCm backend selection, RCCL routing, and native GGUF support, so users do not need to assemble the review branches manually.

git clone --branch feat/rdna4-rocm --single-branch https://github.com/zihaomu/FreeToken.git

See the branch's ROCm source-install instructions, including the tested ROCm 7.14 image and gfx1200/gfx1201 architecture settings.

For upstream review, the same work is being organized as the standalone foundation #132 and Draft follow-ups #133, #134, #135, and #136. Those PRs are review-sized splits; the branch linked above remains the current single-branch version for RDNA4 users.

[zihaomu]

Following the base-chain discussion, I have opened a standalone current-main integration as #132. It preserves the original commit authorship from this PR and from bouclem#1, explicitly credits both contributors, and adds the separately validated RDNA4/ROCm 7.14 work. This removes the cross-fork dependency while giving maintainers a complete RDNA3 + RDNA4 foundation to review.