Draft: Handle vertex AI credentials more gracefully

• Allow them to not be set, in which case we skip init and log a warning
• Support setting credentials in docker-compose

Currently, the model gateway crashes immediately unless GOOGLE_VERTEX_AI_CREDENTIALS is set to a valid credentials file. This puts up an unnecessary barrier to development and receiving contributions. I would like us to move to a development approach that makes this system more accessible to a wider range of contributors. We should not require to have Cloud Provider configs set up when developing locally. Sometimes all you want to test is a change to logging (or something similarly benign), which does not involve talking to Vertex AI or any other remote service.