perf(cuda): extend Q8_1 activation memo to the MMQ MoE path

[dusterbloom]

What

Companion to the mmvq memo (#33): extend the per-eval Q8_1 activation memo to the MMQ MUL_MAT_ID (MoE) path.

At spec-decode verify width (and prefill), Laguna's MoE experts route through MMQ (mmq.cu), which re-quantizes src1 per mul_mat_id call. gate_e and up_e share the same gathered activation + ids, so the same src1→q8_1 is quantized twice per MoE layer × 39 layers every step (reproduced: consecutive quantize calls with identical src1/ids/type/dims).

How

Reuse the existing per-eval ctx.luce_q8_memo (freed LIFO at graph_compute start) for the MMQ ids path — on a hit (up_e after gate_e) reuse gate's q8 buffer and skip the requant. Minimal additions:

• luce_q8_memo_entry gains ids_node/ids_data and an mmq flag — an MMQ block_q8_1_mmq buffer must never cross-reuse an mmvq q8_1 one.
• Keyed on the tensor NODES (src1/ids), since data pointers alias across layers via the graph allocator.
• The q8 buffer is allocated before the transient id buffers, so those sit above it on the strict-LIFO CUDA pool and free first (a held memo buffer must not sit above a buffer that frees before it).

Status

• Env-gated LUCE_MMQ_Q8_MEMO=1 (off by default). Zero effect when off.
• Builds clean; llama-bench pp512 no crash, ~parity (prefill; the real payoff is the verify-width>1 decode path).
• Output verified byte-identical (greedy, off vs on) on the dflash server for this exact logic. Re-confirm on the server after landing, and measure the spec-verify decode delta with the DSpark drafter enabled.

Based on 02cc5286c (the commit the server currently pins via perf/spec-verify-sm86).

---

Summary by cubic

Extend per-eval Q8_1 activation memoization to the MMQ MoE MUL_MAT_ID path to avoid re-quantizing the same gathered activations for gate_e and up_e. Cuts redundant work during spec-decode and prefill with no output changes.

• New Features
  • Reuse ctx.luce_q8_memo in MMQ: keyed by (src1 node, ids node, src0 type, dims); on hit, reuse the q8 buffer and skip requant.
  • Add ids_node, ids_data, and an mmq flag to luce_q8_memo_entry to ensure MMQ (block_q8_1_mmq) never cross-reuses mmvq q8_1.
  • Allocate the q8 buffer before transient id buffers to keep the strict LIFO CUDA pool valid; memo buffers free at graph start.
  • Env-gated via LUCE_MMQ_Q8_MEMO=1 (off by default); no effect when off; outputs validated byte-identical.

^{Written for commit 6c2e686. Summary will update on new commits.}