add ROCm/HIP build instructions for AMD GPUs

[Kaihui-AMD]

Summary

Add build instructions for AMD GPUs via ROCm/HIP. The project already
supports HIP (setup.py detects torch.version.hip and configures HIP
compile flags), but the README only documents a ROCm 6.0 patch without
explaining how to build on AMD hardware.

This addresses #99.

Test environment

• GPU: AMD Radeon R9700 Pro (gfx1201, RDNA 4), 32 GB VRAM
• OS: Ubuntu 24.04, kernel 6.17.0-20-generic
• ROCm: 7.2.3 (container: rocm/pytorch:rocm7.2.3_ubuntu24.04_py3.12_pytorch_release_2.9.1)
• PyTorch: 2.9.1+rocm7.2.3

Build steps

docker run -it --device /dev/kfd --device /dev/dri \
  -e ROCBLAS_USE_HIPBLASLT=0 -e HIP_VISIBLE_DEVICES=0 \
  rocm/pytorch:rocm7.2.3_ubuntu24.04_py3.12_pytorch_release_2.9.1

# inside container
git clone https://github.com/Dao-AILab/causal-conv1d.git
cd causal-conv1d
pip install einops
HIP_ARCHITECTURES=gfx1201 CAUSAL_CONV1D_FORCE_BUILD=TRUE \
  pip install --no-build-isolation -e .

Test results

Functional test:

import torch
from causal_conv1d import causal_conv1d_fn
x = torch.randn(2, 64, 128, device="cuda", dtype=torch.float16)
weight = torch.randn(64, 4, device="cuda", dtype=torch.float16)
out = causal_conv1d_fn(x, weight)
print(out.shape, torch.isnan(out).any(), torch.isinf(out).any())
# torch.Size([2, 64, 128]) tensor(False) tensor(False)

Full test suite:

python3 -m pytest tests/test_causal_conv1d.py -v
================ 9412 passed, 3888 skipped in 166.05s (0:02:46) ======

[Kaihui-AMD]

Hi @tridao @ko3n1g, would you mind taking a look when you get a chance? This is a small README-only change adding ROCm/HIP build instructions — the HIP codepath already works, just undocumented. Thanks!