Decoder adapter caches grow without bound when models are created per request #174

The pydantic and msgspec decoders cache one adapter per model in a plain dict (src/httpware/decoders/_caching.py). The cache holds a strong reference to every model it has seen and never evicts. Code that builds a model per request, for example with pydantic.create_model, grows the cache, and keeps those classes alive, for the life of the decoder.

How to confirm: decode through 10,000 models made with create_model and watch len(decoder._adapters) and the process RSS.

Options if it reproduces: a bounded LRU, or documenting that response models must be long-lived. Either needs a check of what it costs on a cache hit, since the lookup runs on every decoded request.