PyTorch: build flash attention by default, except in CI (#48521)

* PyTorch: build flash attention by default, except in CI

* Variant is boolean, only available when +cuda/+rocm

* desc -> _desc

share/spack/gitlab/cloud_pipelines/stacks/ml-linux-aarch64-cuda/spack.yaml

@@ -12,6 +12,13 @@ spack:
       require: ~cuda
     mpi:
       require: openmpi
+    py-torch:
+      require:
+      - target=aarch64
+      - ~rocm
+      - +cuda
+      - cuda_arch=80
+      - ~flash_attention
 
   specs:
     # Horovod

share/spack/gitlab/cloud_pipelines/stacks/ml-linux-x86_64-cuda/spack.yaml

@@ -12,6 +12,13 @@ spack:
       require: ~cuda
     mpi:
       require: openmpi
+    py-torch:
+      require:
+      - target=x86_64_v3
+      - ~rocm
+      - +cuda
+      - cuda_arch=80
+      - ~flash_attention
 
   specs:
     # Horovod

share/spack/gitlab/cloud_pipelines/stacks/ml-linux-x86_64-rocm/spack.yaml

@@ -11,6 +11,13 @@ spack:
       require: "osmesa"
     mpi:
       require: openmpi
+    py-torch:
+      require:
+      - target=x86_64_v3
+      - ~cuda
+      - +rocm
+      - amdgpu_target=gfx90a
+      - ~flash_attention
 
   specs:
     # Horovod