feat(cli): auto-detect embedding model to enable vector stores and file-search

The letsgo command now detects an embedding model at startup and uses it
to conditionally enable vector stores, inline::file-search, and the
responses API. When no embedding model is found, these features are
disabled with a clear message rather than starting in a broken state.

Users can specify the model explicitly with --default-embedding-model
PROVIDER_ID/MODEL_ID and --default-embedding-dimension DIMENSION. When
omitted, the server scans all active inference providers and selects the
first embedding model found. A --debug flag exposes provider
instantiation logs that are otherwise suppressed during scanning.

Author: mattf