Tracks in-flight streaming inference requests.

Each erllama:infer/4 admit produces a unique reference() and an entry in a public ETS table mapping that reference back to the serving erllama_model gen_statem. erllama:cancel/1 looks up the ref here to find which model owns the request, then casts a {cancel, Ref} event at it.

The table is a fixed-name public ETS so lookups are lock-free from any process. The owning gen_server (this module) is here only to keep the table alive across releases and to clean up entries when a model dies unexpectedly.