[Fwd,SM100,CuTe] Fix split KV OOM with diff headdim + fix SM100 kwarg mismatch

[MatthewBonanni]

Using split KV with diff headdim (e.g. 192/128 for DeepSeek MLA prefill) exceeds SMEM due to float32 partial output doubling the O buffer size. This PR reduces tile size or disables splitting in that case, with a heuristic to ensure optimal performance.

Also fixes a bug introduced in 99d0148 where the SM100 constructor is called with tile_m/tile_n instead of m_block_size/n_block_size, which would cause a TypeError.

This change has been applied to vLLM's FA fork via vllm-project#123 and vllm-project#126 to enable using FA4 for MLA prefill in vLLM (vllm-project/vllm#34732). This PR applies the fix upstream.

Benchmarking performed using vLLM attention benchmark tool:

[tridao]

I feel the right approach is to subtile the O inside the kernel so we can keep the same size of smem. But that's more annoying to impl so for now we can just decrease tile_n.