Feature Request/Bug: Enable native ROCm/AMD RDNA 4 support (GFX1201)

[apollo-mg]

Bug Report Content

Description
The current build system for causal-conv1d assumes an NVIDIA/CUDA environment, blocking native compilation on the
new AMD RDNA 4 architecture (gfx1201 / RX 9070 XT) under ROCm 7.2. While the hardware is capable of running these
operators, the setup.py and C++ source files contain hardcoded CUDA checks.

Environment

• Hardware: AMD Radeon RX 9070 XT (RDNA 4 / gfx1201)
• OS: Ubuntu 22.04 LTS
• ROCm Version: 7.2.0
• PyTorch Version: 2.4.0 (Custom compiled for gfx1201)

Observed Failures

1. Setup Script: setup.py calls torch.cuda.get_device_capability(), which is undefined or errors out on AMD,
   halting the build.
2. Compiler Flags: The build system attempts to pass -gencode flags which are nvcc-specific and unrecognized by
   amdclang++.
3. Source Hardcoding: Files in csrc/ explicitly include <cuda_runtime.h> and use CUDA synchronization primitives
   like __syncthreads() without ROCm/HIP alternatives.

Proposed Solution

• Add ROCm awareness to setup.py by checking torch.version.hip.
• Implement a macro wrapper to conditionally include <hip/hip_runtime.h> when USE_ROCM=1 is set.
• Update the build logic to support amdclang++ compiler flags for the gfx1201 target.

Impact
This blocks the usage of Mamba and Mamba-2 architectures on the latest generation of consumer AMD hardware,
despite the hardware having excellent BF16 and FP8 throughput.

[Kaihui-AMD]

This already works on ROCm/HIP with --no-build-isolation. Tested on gfx1201, all tests pass. Added build instructions in #109.

[apollo-mg]

This already works on ROCm/HIP with --no-build-isolation. Tested on gfx1201, all tests pass. Added build instructions in #109.

Thank you!