trunk/ee2630e634a8d4852cc447e4358aee863241fb85: Skip quantized benchmarks on aarch64 in operator_benchmark (#180532)

The PyTorch development team has resolved a persistent infrastructure issue by merging a targeted code change. The problem was that quantized operator benchmarks—which test the performance of models using reduced precision (like int8) for faster inference—were crashing on ARM-based systems, including Apple's M-series chips and cloud servers. The root cause was the FBGEMM library, a Facebook-developed kernel library for efficient low-precision matrix multiplication, which only supports x86 architectures with AVX2/AVX512 instructions. When PyTorch's CI/CD pipelines tried to run these benchmarks on ARM runners, they failed with an 'unknown architecture' error, breaking the main development branch ('trunk') for over a month.

The fix, authored with assistance from Anthropic's Claude AI and approved by a core maintainer, is elegantly simple: it skips the import and execution of these specific quantized benchmarks (qconv, qlinear, etc.) on non-x86 platforms. This is a pragmatic solution that unblocks development without requiring a full port of the FBGEMM library to ARM. It highlights the growing importance of ARM architecture in the AI ecosystem, from consumer laptops to data center servers, and the ongoing challenges of maintaining cross-platform compatibility for highly optimized, architecture-specific performance libraries.