…_pool flag; unload() only detaches pool

…er, fastapi

…wn cleanup

  - inject shared dependency instance via SuppliedDependency, strict identity match
  vs metadata
  - PreloadedDependencyMismatchError; re-raise in auto-cache path instead of
  swallowing
  - shallow-copy model_init_kwargs per package attempt

  - resolve_shared_base() detects head with supported base dep, negotiates concrete
  package, derives base_key
  - base_key excludes api_key for whitelisted public bases
  - cached metadata fetch; fail-open except auth denials

  - one worker hosts a base model + many heads; head_index slot routing, monotonic
  never-recycled indexes
  - SharedHeadControlPlane load_head/drop_head over req_id-keyed futures; fail on
  timeout and worker death
  - owner backend: eager base load, per-head slot tracker, ref-count teardown,
  in-flight load reservation
  - SharedHeadBackend view: submit_slot/result-callback/drain delegate to owner;
  worker_pid None
  - ModelManager.load_shared_head: create-or-reuse owner under base-key lock,
  empty-owner reaping
  - constructor cleanup on failure; complete in-flight slots on death; stats refresh
  path

  - _load_model_inner resolves shareable head before admission, loads via
  load_shared_head, reuses OOM/evict loop
  - skip admission when base already resident so base VRAM is counted once
  - track head before register_backend so worker death never sees an untracked loaded
  head
  - shared-base worker death marks all hosted heads not-loaded, clears inflight,
  drops backends; heads reload lazily
  - graceful unload/eviction untrack heads; ModelManager.has_shared_base accessor +
  death hook

  - attribute shared owner worker_pid to base_key in stats, so base+heads GPU shows
  once
  - shared head footprint is 0; reclaimable VRAM lives in the base worker
  - eviction works on units (normal models + base_keys); base_key evictable only when
  all heads cold
  - evicting a base_key drops all heads so the worker and its VRAM are actually freed
  - OOM retry evicts the cold unit, never a sibling head that frees nothing
  - ModelManager.shared_owners accessor; drop orphaned _is_still_evictable

This reverts commit 53b0d31.

  - ModelState.pinned; admin/preload T_LOAD pins, even an already-loaded model
  - eviction blocked at _flavor_loaded_idle chokepoint, so all planners honor it
  - a pinned head keeps its shared base resident; explicit T_UNLOAD still unloads
  - no T_LOAD device field: preload and default request both resolve DEFAULT_DEVICE

This reverts commit f017056.