Enable 12 more single-GPU bazel test targets in ROCm CI
- Drop the nine "not yet re-validated" exclusions: api_test, lax_test,
lax_numpy_test, random_lax_test, lax_scipy_sparse_test,
lax_scipy_spectral_dac_test, qdwh_test, shape_poly_test and
export_harnesses_multi_platform_test.
- Drop the three teardown-hang exclusions: cholesky_update_test,
pallas:pallas_vmap_test and pallas:gpu_paged_attention_test. Each passed
five consecutive runs with no hang.
- Brings single-GPU GPU targets from 102 to 114; follows on from #816.
- buffer_callback_test and pallas:ops_test_mgpu stay excluded, for the
py_import incompatibility and Mosaic GPU respectively.
The complex-GEMM gap behind most of these exclusions no longer reproduces:
C64 and C128 matmul return correct results on ROCm 7.14. Validated on
MI300X (gfx942); the bazel CI pool is gfx950, so CI is the real gate.
