Add support for gfx12

[mgehre-amd]

Motivation

Avoiding vLLM to fail with

File "/…/.venv/lib/python3.12/site-packages/aiter/__init__.py", line 50, in <module>
    from .jit import core as core
  File "/…/.venv/lib/python3.12/site-packages/aiter/jit/core.py", line 260, in <module>
    gfx = get_gfx()
          ^^^^^^^^^
  File "/…/.venv/lib/python3.12/site-packages/aiter/jit/utils/chip_info.py", line 86, in get_gfx
    gfx_num = get_gfx_custom_op()
              ^^^^^^^^^^^^^^^^^^^
  File "/…/.venv/lib/python3.12/site-packages/aiter/jit/utils/torch_guard.py", line 281, in wrapper_custom
    else getattr(torch.ops.aiter, f"{loadName}")(
         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/…/.venv/lib/python3.12/site-packages/torch/_ops.py", line 1255, in __call__
    return self._op(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/…/.venv/lib/python3.12/site-packages/aiter/jit/utils/torch_guard.py", line 321, in outer_wrapper_dummy
    else (torch.empty(1, device=device), wrapper(*args, **kwargs))
                                         ^^^^^^^^^^^^^^^^^^^^^^^^
  File "/…/.venv/lib/python3.12/site-packages/aiter/jit/utils/torch_guard.py", line 197, in wrapper
    return func(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^
  File "/…/.venv/lib/python3.12/site-packages/aiter/jit/utils/chip_info.py", line 46, in get_gfx_custom_op
    return get_gfx_custom_op_core()
           ^^^^^^^^^^^^^^^^^^^^^^^^
  File "/…/.venv/lib/python3.12/site-packages/aiter/jit/utils/chip_info.py", line 72, in get_gfx_custom_op_core
    raise RuntimeError(f"Get GPU arch from rocminfo failed {str(e)}")

when importing aiter. Maybe we should avoid maintaining an explicit list of GPUs in AITER?

Technical Details

Added gfx12 wherever other gfx targets are mentioned.

Test Plan

None

Test Result

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[schung-amd]

Verified that this resolves ROCm/legacy-rocm-build#5656. Tested with two gfx1201 nodes using the rocm/vllm:latest Docker image as a base and installing an AITER wheel with this PR on top, was able to perform distributed inference with ray after the patch.