cuDNN SDPA executor does not support batched broadcast across attention heads mask

[IvanYashchuk]

🐛 Bug

To reproduce:

1. Pull the branch of Update cuDNN SDPA execution functions to use the strides of attn_mask #1798 if not merged yet.
2. Run the test added in Update cuDNN SDPA execution functions to use the strides of attn_mask #1798 (pytest thunder/tests/test_cudnn_executor.py::test_make_cudnn_sdpa_backward_graph_with_mask):
   

   lightning-thunder/thunder/tests/test_cudnn_executor.py

   Lines 65 to 72 in 839d65f

   	# "B1TS" case is not working
   	mask = torch.ones(batch_size, 1, sequence_length, sequence_length, dtype=torch.bool, device="cuda")
   	mask = torch.where(mask, 0.0, float("-inf")).to(torch.bfloat16)
   	# Reason: dbias cannot be reduced just across heads.
   	# Supported reduction patterns are [1,1,T,S], [1,H,T,S], [B,H,T,S] at: (!batch_dim_reduction_requested) && head_dim_reduction_requested
   	# See issue: ...
   	with pytest.raises(cudnn.cudnnGraphNotSupportedError, match="No valid engine configs"):
   	_make_cudnn_sdpa_backward_graph(q, k, v, mask, 0.0, None)

Supporting this case is required for batched packed sequences (#1758).