perf: frank-default-small ctx 200000/stream, --no-mmap, batch 3072/1536

Lower each stream to 200000 ctx (pool 600000) to free KV VRAM for the larger
batch. Add --no-mmap so the CPU-resident MoE experts load eagerly into RAM
(avoids decode page faults). Keep KV on GPU. Experts already fully on CPU.

scripts/serve-llama-frank-small.sh

@@ -35,6 +35,7 @@ fi
 : "${BATCH_SIZE:=2048}"
 : "${UBATCH_SIZE:=512}"
 : "${CTX_CHECKPOINTS:=0}"
+: "${NO_MMAP:=0}"
 : "${THREADS:=8}"        # 7700X physical cores
 : "${THREADS_HTTP:=8}"
 
@@ -66,6 +67,7 @@ exec "$LLAMA_SERVER" \
     --batch-size "$BATCH_SIZE" \
     --ubatch-size "$UBATCH_SIZE" \
     --ctx-checkpoints "$CTX_CHECKPOINTS" \
+    $([ "$NO_MMAP" = "1" ] && printf -- "--no-mmap") \
     --threads "$THREADS" \
     --threads-http "$THREADS_HTTP" \
     --alias "$ALIAS" \