fix(rocm): make TVM-FFI index and store JIT kernels portable to HIP

[zihaomu]

Draft follow-up to #132. This branch is based on the #132 head. Until #132 merges, GitHub Files changed also includes the foundation diff; the incremental review scope is the four files listed below.

Summary

• disable NVIDIA programmatic-dependent-launch instructions and launch attributes on ROCm while preserving the CUDA path;
• accept kDLROCM tensors in the generic index and store TVM-FFI kernels;
• keep the CUDA PTX implementation unchanged and use the HIP-compatible common launch path;
• use kWarpThreads as the single source for wave-size assertions;
• cover split=1/2/4 row copies, masked vocabulary misses, both index widths, and both store index widths.

Incremental review scope

Relative to zihaomu:feat/rocm-rdna3-rdna4-foundation:

• python/freetoken/kernel/csrc/include/freetoken/utils.cuh
• python/freetoken/kernel/csrc/jit/index.cu
• python/freetoken/kernel/csrc/jit/store.cu
• tests/kernels/test_jit_index_store.py

Diff size: 88 additions, 14 deletions.

Dependency

• Depends on feat(rocm): add RDNA3 and RDNA4 runtime foundation #132.
• Does not include optional-backend gating, RCCL routing, or GGUF support.

Validation

• GPU: AMD Radeon AI PRO R9700 (gfx1201)
• PyTorch 2.11.0+rocm7.14.0, HIP 7.14.60850
• cold JIT plus pinned-memory focused run: 14 passed
• warm tests/kernels/test_jit_index_store.py: 6 passed
• covered fp32 D=256 (split=2), D=512 (split=4), lower/upper masked reset, int64 indexing, and int32/int64 store indices.

Compatibility

CUDA branches remain behind the non-ROCm preprocessor path. Physical NVIDIA regression testing is not claimed; this PR remains Draft until #132 and CUDA CI establish the final base.