[ROCm] Support the INT4 QAT kernel on ROCm

[LZ-QWQ]

The int4_qat kernel does not build on a ROCm PyTorch install, so the INT4 QAT path is unusable on AMD GPUs. This fixes the two blockers (a HIP warp-shuffle mask and a const_data_ptr symbol mangling mismatch under hipcc).

Build

• CUDA: unchanged. Every modification is behind #if defined(__HIP_PLATFORM_AMD__) or torch.version.hip is not None, so the CUDA compile path is byte-for-byte identical.
• ROCm: pip install . just works. The gfx target comes from PYTORCH_ROCM_ARCH (defaults to gfx950), since hipcc does not accept the nvcc-only flags.

Verification (MI355X / ROCm 7.2, CUDA reference from an RTX 4090)

• 144 cases covering the block shapes, tensor shapes and dtypes the kernel dispatches on: outputs are bit-exact against CUDA.
• E2E tools/convert_hf_to_int4_direct.py on Qwen3-0.6B: 196 layers converted, dequantization error within the expected range for group-wise INT4 (worst 0.1051).