[lowering] Support SDPA where the value head dim differs from query/key

[gokulkrishna98]

scaled_dot_product_attention lowering assumed the output had the same type as query (D_v == D_k) and asserted on it, so attention with a differing value head dim failed to convert even though eager PyTorch handles it.

• Derive the result type from the value head dim ([B, n_q_heads, T_q, D_v]) in _sdpa_decompose and in all three rank paths of replace_sdpa.
• Emit the composite op only when D_v is provably equal to D_k; its interface cannot represent another output head dim, so D_v != D_k and dynamic head dims lower the decomposition directly.
• The rank > 4 restore path rebuilt the shape from the query, which carries D_k; it now uses the query's leading dims with the value head dim.
• The default 1/sqrt(D_k) scale still uses the query head dim.

Repro

class SDPA(nn.Module):
    def forward(self, q, k, v):
        return F.scaled_dot_product_attention(q, k, v, enable_gqa=True)

# q: [8, 4, 1, 32], k: [8, 1, 36, 32], v: [8, 1, 36, 24]

Asserted before this change; converts now.

Tests

• TestSDPA::test_sdpa_value_head_dim_differs — rank 4 and rank 6, causal / explicit-mask / maskless, static and dynamic, GQA.
• TestSDPA::test_sdpa_rank3_value_head_dim_differs, TestSDPA::test_sdpa_dynamic_head_dim.
• IR tests asserting the composite is emitted for D_v == D_k and that D_v != D_k and dynamic head dims lower inline.

pytest tests/ops/test_ops.py -k SDPA tests/ops/test_ops_ir.py tests/composite_ops/test_sdpa.py — 2083 passed.