E2: MultiLinear (fwd_multi_linear) primitive for MLA

[tlkahn]

MultiLinear (fwd_multi_linear)

Per-head batched matmul primitive required by MLA attention.

Spec: doc/plans/engine-inference-core.md § E2

Part of: #105 (E2) · Epic: #56

Tasks

• fwd_multi_linear: weight shape [H, O, I], per-head batched matmul
• Quantized variant via mlx_quantized_matmul with per-head scales/biases
• Port semantics from mlx-lm mla.py MultiLinear / QuantizedMultiLinear
• Unit tests on synthetic weights (bf16 + quantized)

Notes

• Consumed by MLA as embed_q and unembed_out