fix[next]: Detect ROCm device architecture instead of hard-coding gfx942

[tehrengruber]

Problem

_query_device_arch() in otf/compilation/common.py returns a hard-coded "gfx942" (MI300A) for ROCm devices — an old TODO. Every GPU binary is therefore compiled for gfx942 only (get_device_arch() feeds both the DaCe and the gtfn/CMake build); on any other AMD GPU the fat binary contains no matching code object and the first kernel launch segfaults inside libamdhip64 (the stub-based hipLaunchKernel does not return a clean error for this case). Observed on a gfx1103 APU; setting HIPARCHS explicitly works around it.

Fix

Query the architecture from the device via cupy.cuda.runtime.getDeviceProperties(0)["gcnArchName"], stripping the target-feature suffix (e.g. gfx90a:sramecc+:xnack- → gfx90a), with the same error handling as the CUDA branch. The CUDAARCHS/HIPARCHS environment override precedence is unchanged. Verified on a gfx1103 APU (ROCm 7.2): detection returns gfx1103.

AI disclaimer: Description and code were developed with the help of LLMs. Code was reviewed in detail, description briefly read, too verbose, but doesn't matter.