Set PyTorch parallelism limits for aarch64/{generic,neoverse_n1,neoverse_v1}

[bedroge]

Still keeping this as draft to potentially solve other issues.

[boegel]

You're seeing out-of-memory errors only for aarch64/* targets?

That's interesting, all partitions in our main build cluster have name amount of RAM per core...

[bedroge]

Not completely sure yet, but it definitely worked fine for skylake_av512 and zen5. It failed for neoverse v1 with:

cc1plus: out of memory allocating 65536 bytes after a total of 319553536 bytes
cc1plus: out of memory allocating 139696 bytes after a total of 301793280 bytes
cc1plus: out of memory allocating 3201600 bytes after a total of 98828288 bytes
cc1plus: out of memory allocating 65536 bytes after a total of 348848128 bytes
virtual memory exhausted: Cannot allocate memory
cc1plus: out of memory allocating 193048 bytes after a total of 369950720 bytes
cc1plus: out of memory allocating 145472 bytes after a total of 401604608 bytes
cc1plus: out of memory allocating 146840 bytes after a total of 386990080 bytes
cc1plus: out of memory allocating 65536 bytes after a total of 548405248 bytes
cc1plus: out of memory allocating 151096 bytes after a total of 391184384 bytes
cc1plus: out of memory allocating 65536 bytes after a total of 620494848 bytes

Started builds for the other Arm targets now (the updated hooks file is not applied yet anyway, I just noticed, because it's still using the one from CVMFS ).