Various problems on arm64

[omegacoleman]

I'm testing FBGEMM on a Macbook M1. The kernels are fine, but the build system and ISA detection fail. Currently

1. cpuinfo impl, from our dependencies, is in an early stage on arm64. No SVE detection is performed on macOS, and on Linux, it depends on the kernel.
2. What's worse, the env override is missing on arm64
3. On our side, the ISA selector function and the actual kernels are built with different macro definitions.
4. A few misalignments in the build script also create problems.
5. The SVE kernel currently produces wrong results

This issue is created to track those problems.