trunk/4eb3032930d501edc0c4accef34e6936be0b3e01: Fix flex decoding CPU test discovery on CUDA hosts (#185380)

test_flex_decoding.py picked one global test_device at import time. On CUDA-capable hosts this became only cuda, so instantiate_device_type_tests never created TestFlexDecodingCPU and CI's CPU copy-paste test commands could not resolve the generated class.

Instantiate CPU flex-decoding tests alongside CUDA tests when the CPU flex-attention platform is supported. The dtype coverage and CPU-only skips are now selected per instantiated device, and helper allocations use the active device rather than the old global device choice. This keeps the existing generated test-name suffixes intact so targeted unittest commands continue to work.

I considered changing common_device_type discovery or relying on PYTORCH_TESTING_DEVICE_ONLY_FOR, but this test file was the source of the single-device filtering and already has per-backend flex-attention support predicates. Keeping the fix local avoids changing global test infrastructure behavior.

Fixes

Approved by: https://github.com/drisspg