python3Packages.torch: Disable MKLDNN on aarch64-darwin

It negatively impacts performance, this is also what official pytorch
build does.

In my test in a Macbook Pro M1 2020, using [pianotrans][1] to transcribe
[cut_liszt.opus][2], transcription time:

Enable mkldnn and disable Accelerate.framework: ~88s

Disable mkldnn and disable Accelerate.framework: ~21s

Disable mkldnn and enable Accelerate.framework: ~9s

The final result is close to using torch-bin.

See also https://github.com/NixOS/nixpkgs/pull/219104

[1]: https://github.com/azuwis/pianotrans
[2]: https://github.com/azuwis/pianotrans/raw/master/test/cut_liszt.opus

pkgs/development/python-modules/torch/default.nix

@@ -1,6 +1,6 @@
 { stdenv, lib, fetchFromGitHub, fetchpatch, buildPythonPackage, python,
   cudaSupport ? false, cudaPackages, magma,
-  mklDnnSupport ? true, useSystemNccl ? true,
+  useSystemNccl ? true,
   MPISupport ? false, mpi,
   buildDocs ? false,
 
@@ -17,6 +17,10 @@
   # Unit tests
   hypothesis, psutil,
 
+  # Disable MKLDNN on aarch64-darwin, it negatively impacts performance,
+  # this is also what official pytorch build does
+  mklDnnSupport ? !(stdenv.isDarwin && stdenv.isAarch64),
+
   # virtual pkg that consistently instantiates blas across nixpkgs
   # See https://github.com/NixOS/nixpkgs/pull/83888
   blas,