Release 1.1.0

Adds a time-to-live to the per-worker inference cache to bound how long
stale entries can sit in memory between requests, in addition to the
existing size cap.

• Replaced functools.lru_cache in
  api/services/classification_service.py with
  cachetools.TTLCache(maxsize=10_000, ttl=3600) + @cached. Entries
  now expire one hour after insertion, in addition to being evicted
  when the size cap is reached.
• A threading.Lock is wired in via the lock= argument of @cached.
  The lock is held only around cache reads and writes, not around the
  wrapped inference call, so concurrent requests still run in parallel.
• Expiry is lazy: stale entries are dropped when their key is next
  accessed or when an insertion scans the cache. There is no background
  sweep, so cache memory is reclaimed on use rather than on a timer.
• Added cachetools (>=5.3.0,<6.0.0) as a runtime dependency.

Note: if a worker is still being OOM-killed after this change, the cache
is unlikely to be the cause — at 10,000 entries × ~1–2 KB it is bounded
at roughly 10–20 MB. The more common culprits under sustained
transformers load are PyTorch allocator fragmentation and thread-stack
growth; mitigations there include gunicorn's --max-requests to
recycle workers periodically, restricting OMP_NUM_THREADS, and
tuning MALLOC_TRIM_THRESHOLD_.