fix(dsa): fall back for unsupported odd cuDNN top-k

[JackRao123]

Summary

• route odd indexer top-K selection through masked torch.topk to avoid cuDNN Frontend 1.26.0's vector-width assertion
• retain cuDNN score generation and FlashMLA sparse attention, and keep even K on the fused cuDNN selector
• cover the CP32 regression shape, varlen masking, score returns, and even-K dispatch

Test plan

• Run Ruff checks on the changed files
• Compile both changed Python files
• Focused pytest passes on the B200 devbox (4 passed)
• Exact [235, 705], top_k=705 patched primitive passes on B200
• Repeat the exact CP32 15,000-token trainer forward after building the patched image

[pstefa1707]

This will b v slow if its raw pytorch right? When do we hit this case?
Should we have a triton kernel to cover it?