fix(vram): stop coexisting pipelines from rebuilding every run

BuildPipeline.build() called comfy's unload_all_models() to free
ComfyUI-native torch weights, but that blanket free_memory(1e30) went
through our own free_memory hook and was read as "destroy all QuantFunc
pipelines" — so running a second pipeline destroyed the first (and
vice-versa), forcing a full rebuild on every repeat run even with
activate_unload on.

- BuildPipeline self-cleanup now routes through free_comfy_native_models(),
  which calls the captured ORIGINAL (un-hooked) free_memory: clears
  comfy-native torch only, never touches QuantFunc pipelines.
- The free_memory hook is now a measured graded ladder: VRAM pressure ->
  offload GPU->CPU only (offload frees the GPU; destroying an already-
  offloaded pipeline frees zero VRAM, so VRAM-triggered destroy was pure
  friendly-fire). RAM-low -> disk-page backups (madvise). Destroy is
  reserved for genuine host-RAM exhaustion only, clamped so the reversible
  disk-page rung always precedes the destructive rung.
- LRU bookkeeping + resident-pipeline cap (QUANTFUNC_MAX_RESIDENT_PIPELINES,
  default 3) to bound growth now that blanket no longer destroys.

Verified live (device 1, RTX 3060): Z-Image + Qwen-Edit coexisting,
repeat runs reuse both pipelines (~22s) instead of rebuilding (64s+),
with no friendly-fire destroy.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>