Migrate MacOs wheel binary builds to ephemeral M1 runners (#110432)

Surprisingly there are no speed difference between running the cross-compilation on `macos12-xl` (x86_64 12 core machine) and `macos-13-xlarge` (m1 6 core machine)

Most of the changes are on the https://github.com/pytorch/builder side:
- 50a6e91f97 skips installing mkl on M1 machines
- bbb29b0467 same for llvm-9
- 8bcc83dbb1 bumps minimal numpy version to 1.19 (as 1.17 is not available for m1)
- cc4f1f9055 skips building tests/distributed for M1
Pull Request resolved: https://github.com/pytorch/pytorch/pull/110432
Approved by: https://github.com/kit1980

.github/scripts/generate_ci_workflows.py

@@ -60,7 +60,7 @@ class BinaryBuildWorkflow:
     branches: str = "nightly"
     # Mainly for macos
     cross_compile_arm64: bool = False
-    xcode_version: str = ""
+    macos_runner: str = "macos-12-xl"
 
     def __post_init__(self) -> None:
         if self.abi_version:
@@ -307,7 +307,8 @@ MACOS_BINARY_BUILD_WORKFLOWS = [
         build_configs=generate_binary_build_matrix.generate_wheels_matrix(
             OperatingSystem.MACOS_ARM64
         ),
-        cross_compile_arm64=True,
+        cross_compile_arm64=False,
+        macos_runner="macos-13-xlarge",
         ciflow_config=CIFlowConfig(
             labels={LABEL_CIFLOW_BINARIES, LABEL_CIFLOW_BINARIES_WHEEL},
             isolated_workflow=True,