[rocm7.0_internal_testing] mx fpx: Skip test_blockwise_mxfloatx (#2439)

Issue is tracked by SWDEV-535267

This PR skips two test methods
(test_blockwise_mxfp8_nvfp4_error_messages and
test_blockwise_nvfp4_compile) when running on ROCm platform to address
an issue tracked by SWDEV-535267.

- Adds @skipIfRocm decorators to skip tests that are failing on ROCm 7.0
internal testing
- Targets specific matrix multiplication tests related to MX float
formats

Signed-off-by: Jagadish Krishnamoorthy <jagadish.krishnamoorthy@amd.com>

Author: jagadish-amd

test/test_matmul_cuda.py

@@ -1385,6 +1385,7 @@ def test_blockwise_mxfp8_nvfp4_mxfp4_numerics(self, test_case_name, fast_accum,
             sqnr = compute_error(C_ref, C)
             assert sqnr.item() > approx_match_sqnr_target
 
+    @skipIfRocm
     @unittest.skipIf(not PLATFORM_SUPPORTS_FP8 or IS_WINDOWS, f8_msg)
     @parametrize("recipe", ["mxfp8", "nvfp4"])
     def test_blockwise_mxfp8_nvfp4_error_messages(self, device, recipe) -> None:
@@ -1618,6 +1619,7 @@ def test_blockwise_mxfp8_compile(self) -> None:
         )
         torch.testing.assert_close(C, C_ref, atol=0, rtol=0)
 
+    @skipIfRocm
     @unittest.skipIf(not PLATFORM_SUPPORTS_MX_GEMM, mx_skip_msg)
     def test_blockwise_nvfp4_compile(self) -> None: