feat(batch): heterogeneous scheduler — PagedKVAllocator + GdnStateSlabAllocator (ADR-063)

[ohdearquant]

Continuous batching for the hybrid model needs two allocators (implements part of ADR-063 serving): PagedKVAllocator (6 GQA layers, context-linear, quantizable pages) + GdnStateSlabAllocator (18 layers, fixed-size per sequence, checkpointable). Admission control must account for both.

Tasks

• PagedKVAllocator (fixed pages, block tables, free queues; bytes/token = 6·2·512·dtype) + GdnStateSlabAllocator (1 slab/seq)
• admission: free_kv_pages ≥ ceil((P+M)/T) AND free_gdn_slabs ≥ 1 AND scratch fits AND adapter-generation compatible; high-watermark (0.85) soft-reservation for interactive, hard for bench
• eviction/preemption: LRU prefix pages → pause → PreemptedRecompute (free KV+GDN, retain CPU history); never evict active KV without recompute
• decode batch mixes GQA (M=batch GEMV) with GDN (per-seq sequential recurrence)

Acceptance (ADR-064 gates)

• allocator NEVER allocates KV for GDN layers (runtime assertion feat(kv_cache): runtime KV-layout assertion (6 GQA layers, fail-closed) #170)
• no OOM in hard-reservation bench; serving batch tok/s no >7% regression
• quantized KV (from feat(kv_cache): quantized KV cache storage (int8/int4/fp8) #118) measurably raises max concurrency

Ref: d5§7, ADR-063. Study vLLM Hybrid KV Cache Manager + PagedAttention.