[aarch64] Pin ACl to v22.11

[atalman]

This switches ACL back to working configuration.

At a time of this PR:
#1370

This code was added in Compute Library v23.05 :

if 'v8a' in env['arch']:
        print("INFO: multi_isa armv8-a architecture build doesn't enable __ARM_FEATURE_FP16_VECTOR_ARITHMETIC. Use armv8.2-a or beyond to enable FP16 vector arithmetic support")
        env.Append(CXXFLAGS = ['-march=armv8-a']) # note: this will disable fp16 extension __ARM_FEATURE_FP16_VECTOR_ARITHMETIC

Ref:
ARM-software/ComputeLibrary@6c713f0#diff-d78c5c377c03dd251edc72cdb0ae7d9986fa08361eb392187218f8c670c0bc4eR311

[malfet]

Sure, those I've tried this one few weeks back and it did not work

This is why I've collected the backtrace in pytorch/pytorch#109312 to show that it comes from https://github.com/oneapi-src/oneDNN/blob/99c3fb44488922554696276dc88e8b6279eb662d/src/common/utils.cpp#L113

[snadampal]

@malfet , it's crashing even without ACL. I have built the wheels with USE_MKLDNN=ON but USE_MKLDNN_ACL=OFF, and observed the crash. So, i think it's the onednn build we need to check.

[snadampal]

@malfet , please don't revert the ACL version. Please check my comment on the original issue
pytorch/pytorch#109312 (comment)

[malfet]

@malfet , it's crashing even without ACL. I have built the wheels with USE_MKLDNN=ON but USE_MKLDNN_ACL=OFF, and observed the crash. So, i think it's the onednn build we need to check.

@snadampal sorry, I should have made it clear in the issue, that this is not an ACL problem, but rather a oneDNN one

[malfet]

No longer needed