Regressions on "PyTorch Flex Decoding, PVC (Max 1100), pinned"

[anmyachev]

CI: https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/29408876154

Logs:

The following tests failed consistently: ['test/inductor/test_flex_decoding.py::TestFlexDecodingXPU::test_captured_buffers_all_dims_bfloat16_xpu_bfloat16', 'test/inductor/test_flex_decoding.py::TestFlexDecodingXPU::test_captured_buffers_all_dims_float16_xpu_float16', 'test/inductor/test_flex_decoding.py::TestFlexDecodingXPU::test_captured_buffers_all_dims_float32_xpu_float32', 'test/inductor/test_flex_decoding.py::TestFlexDecodingXPU::test_do_not_trigger_dynamic_shapes_on_empty_block_mask_xpu', 'test/inductor/test_flex_decoding.py::TestFlexDecodingXPU::test_max_autotune_with_captured_xpu', 'test/inductor/test_flex_decoding.py::TestFlexDecodingXPU::test_non_sparse_mulitple_block_size_xpu']