v0.3.0

v0.3.0 — model auto-download, and a GPU package that actually works

Missing models are now fetched, SHA-256 verified against a manifest
compiled into the binary, and cached per platform. The URL is pinned to an
immutable release tag, so it is a version rather than a moving target.

Also fixes a bug present since the first release: the packaged binary
shipped no onnxruntime_providers_shared library, so --cuda and --tensorrt
could not load from a release archive on either platform.