Live cache rebuild grows MoE before shrinking KV/mamba, so a resize that passes the fit check can OOM #643

Engine.rebuild_runtime_cache (engine/engine.py:1006-1021) always resizes MoE, then KV, then mamba, and each pool frees only its own old tensors. A live resize that moves memory from KV or mamba into MoE therefore allocates the larger MoE cache while the old KV/mamba pools are still resident. The fit check (validate_rebuild, engine.py:972) prices only the target geometry, so on a nearly full card a resize that passes it can OOM and roll back. (A separate rollback failure after such an OOM is #526.)

Reproduce: serve an offloaded-MoE model near full, then ft ctl cache --moe <larger> --kv <smaller> with a target total at or below the current one. A CPU check that drives rebuild_runtime_cache with fake pools under a shared cap (current moe=40, kv=50; target moe=60, kv=30; cap 100) fails on the MoE allocation while kv=50 is still allocated.

Expected: apply the shrinking resizes before the growing ones, so any geometry that passes the fit check is reachable live.

Environment: main 07813248 (source build), RTX 5080 16 GB (sm_120), driver 610.57.04, Ubuntu 24.04, Python 3.14, torch 2.11.0+cu130, triton 3.6.0; checkpoint nvidia/Qwen3.8-Flash-Next-NVFP4.

Implementation (on my fork, with a test): 50caea2. Fork PR: gdevenyi#22. #650 now fixes this upstream.