[Examples] Gate CUDA-only and flash_attn-dependent examples
Three example tests cannot run on non-CUDA backends:
* `test_example_gemm_intrinsics` lowers `tl.ptx_ldmatrix`, which is
NVIDIA PTX-only.
* `test_example_chunk_delta_bwd_compilation` uses `T.tfloat32`, which has
no CDNA lowering.
* the two `sparse_gqa_decode_varlen_*` blocksparse examples import
`flash_attn`, which is installed by requirements-test-cuda.txt but has
no equivalent in requirements-test-rocm.txt.
Gate the first two with `requires_cuda` and the blocksparse pair with
`requires_package("flash_attn")` so they keep running on CUDA and skip
cleanly elsewhere.
andyluo7
3f2fa07eb3