Port of optiq fused_quant_sdpa._prefill_flashattn_n_tiled: a
FlashAttention-2 loop over the KV N axis with mx.quantized_matmul as
the inner kernel. Never materializes the full [..., L, N] scores
matrix — the per-tile transient is bounded by FUSED_N_CHUNK, which is
the whole point (stock u4 KV prefill peaks ABOVE fp16 KV at long
context; see the oracle's module docstring). Op composition order
mirrors the oracle exactly: parity is tier a (bit-exact) vs the python
fused path; vs quantizedSdpaUnfused it is tier b by construction
(online softmax ≠ one-shot precise softmax in bf16).
Port of optiq fused_quant_sdpa._prefill_flashattn_n_tiled: a FlashAttention-2 loop over the KV N axis with mx.quantized_matmul as the inner kernel. Never materializes the full [..., L, N] scores matrix — the per-tile transient is bounded by FUSED_N_CHUNK, which is the whole point (stock u4 KV prefill peaks ABOVE fp16 KV at long context; see the oracle's module docstring). Op composition order mirrors the oracle exactly: parity is tier a (bit-exact) vs the python fused path; vs quantizedSdpaUnfused it is tier b by construction (online softmax ≠ one-shot precise softmax in bf16).