Introduce rocm-only tag and remove if_rocm_is_configured

This replaces all `if_rocm_is_configured` guards in `stream_executor/rocm/...` with a filtering tag `rocm-only`. The CUDA build on the CI gets adjusted to skip those targets.

This uncovered some additional problems that get fixed as well:

- A wrong library name for the hipfft library in the Bazel CUDA configuration
- A wrong test case in the RocmVersionParser test that so far has not been running anywhere.
- Missing tags for the platform alias targets in `stream_executor/BUILD`

PiperOrigin-RevId: 678548961

third_party/gpus/rocm_configure.bzl

@@ -476,7 +476,7 @@ def _create_dummy_repository(repository_ctx):
             "%{hipblas_lib}": _lib_name("hipblas"),
             "%{miopen_lib}": _lib_name("miopen"),
             "%{rccl_lib}": _lib_name("rccl"),
-            "%{hipfft_or_rocfft}": _lib_name("hipfft"),
+            "%{hipfft_or_rocfft}": "hipfft",
             "%{hipfft_or_rocfft_lib}": _lib_name("hipfft"),
             "%{hiprand_lib}": _lib_name("hiprand"),
             "%{hipsparse_lib}": _lib_name("hipsparse"),