[wave] NSA: compressed attention forward kernel

## Parent
Part of #1243 — DeepSeek NSA kernels for MI350

## Description

Implement the forward pass of the compressed attention branch. This computes standard causal attention over the mean-pooled (compressed) KV representations, producing both the output and log-sum-exp (LSE) values needed by the top-k selection stage.

### Operation

```
Input: Q [B, M, H, D], K_cmp [B, N//bs, G, D], V_cmp [B, N//bs, G, D]
Output: O_cmp [B, M, H, D], LSE_cmp [B, H, M]
```

Causal masking: `q_idx <= (kv_idx + 1) * block_size - 1`

### Requirements

- GQA-compatible (G KV groups, H query heads)
- Return LSE alongside output (needed for top-k scoring in selection branch)
- Causal block mask aligned to compression block boundaries
- FP16 matmuls, FP32 softmax accumulation
- Must produce numerically stable LSE values

### MI350 considerations
- The compressed sequence length is N/block_size (e.g., 1024 for 64k/64), making this a relatively small attention — optimize for this regime
- Consider using matrix cores for the QK^T and PV matmuls even at this reduced size
- May be able to fuse with the mean pooling from the compression kernel

### Depends on
- #1244 (design doc)
- #1245 (token compression kernel)

### References
- `compression_attention` in tilde-research/nsa-impl uses FlexAttention
- NSA paper Section 3.1