Track and follow MI355 unit test fails in upstream trunk.

[anatoliylitv]

Upstream Pytorch repo have few MI355 units fails which blocks development. Need to track create items and fix them.
Also need to sync with core unit test board.

[anatoliylitv]

https://ontrack-internal.amd.com/browse/SWDEV-571650

[ROCm QA] On MI355 Test test_effn_attn_bias_padding_misaligned failing due to incorrect result

PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_cuda_repro.py CudaReproTests.test_effn_attn_bias_padding_misaligned

[anatoliylitv]

https://ontrack-internal.amd.com/browse/SWDEV-571645

[ROCm QA] On MI355 Test test_addmm_relu_tunableop_rocm_cuda_float64 failing due to an illegal memory access

2025-12-09T22:10:17.8116511Z ----------------------------- Captured stderr call -----------------------------
2025-12-09T22:10:17.8116824Z /var/lib/jenkins/pytorch/test/test_linalg.py:7393: UserWarning: HIP warning: an illegal memory access was encountered (Triggered internally at /var/lib/jenkins/workspace/aten/src/ATen/hip/impl/HIPGuardImplMasqueradingAsCUDA.h:83.)
2025-12-09T22:10:17.8116877Z res1 = f(t, m, v, alpha=alpha, beta=beta)

PYTORCH_TEST_WITH_ROCM=1 python test/test_linalg.py TestLinalgCUDA.test_addmm_relu_tunableop_rocm_cuda_float64

[anatoliylitv]

PRS to fix unit test fail:
pytorch#169694
and to address discussion:
pytorch#170306

rocm-mi355 / linux-noble-rocm-py3.12-mi355 / test (default, 1, 6, linux.rocm.gpu.mi355.1) (gh)
test/test_transformers.py::TestSDPACudaOnlyCUDA::test_flash_attention_vs_math_ref_grads_batch_size_8_seq_len_q_2048_seq_len_k_579_head_dim_256_is_causal_False_dropout_p_0_48_float16_scale0_enable_gqa_True_n_heads1_cuda_float16