Disable AOTriton for gfx906 and gfx908 in PyTorch builds (#3759)

This PR adds gfx906 and gfx908 to AOTRITON_UNSUPPORTED_ARCHS in the
PyTorch build script.

AOTriton is not currently supported on these architectures, which can
lead to build failures when attempting to enable Flash Attention.
Disabling it ensures PyTorch wheels build successfully for these
targets.

This aligns the build configuration with the current AOTriton
architecture support.

---------

Co-authored-by: Aravind Ravikumar <[email protected]>

Author: araravik-psd

external-builds/pytorch/build_prod_wheels.py

@@ -783,8 +783,8 @@ def do_build_pytorch(
     is_pytorch_2_11_or_later = pytorch_build_version_parsed.release[:2] >= (2, 11)
 
     # aotriton is not supported on certain architectures yet.
-    # gfx101X/gfx103X: https://github.com/ROCm/TheRock/issues/1925
-    AOTRITON_UNSUPPORTED_ARCHS = ["gfx101", "gfx103"]
+    # gfx906/gfx908/gfx101X/gfx103X: https://github.com/ROCm/TheRock/issues/1925
+    AOTRITON_UNSUPPORTED_ARCHS = ["gfx906", "gfx908", "gfx101", "gfx103"]
     # gfx1152/53: supported in aotriton 0.11.2b+ (https://github.com/ROCm/aotriton/pull/142),
     #   which is pinned by pytorch >= 2.11. Older versions don't include it.
     if not is_pytorch_2_11_or_later: