feat: use anthropic on vertex if configured

feat: use anthropic on vertex if configured

To do this, it moves the creation of a ChatModel into a the llm_factory.

When we find the DUO_WORKFLOW__VERTEX_PROJECT_ID we'll try to connect to Vertex-Anthropic. If it's not set, but ANTHROPIC_API_KEY is set, we'll connect to regular-Anthropic. This means that we can gradually transition to using Vertex-Anthropic in all environments.

When DUO_WORKFLOW__VERTEX_PROJECT_ID is set, we also require DUO_WORKFLOW_SERVICE__VERTEX_LOCATION to be set. For now, this is because we need to hardcode connecting to us-east5 for Vertex-Anthropic, the models we need are not available in us-east1. We'll need to change this behaviour based on RUNWAY_REGION if we want to deploy this service in multiple regions.

During the initialization of the server, we also explicitly test if the model is accessible with the configured environment variables. This replaces the startup probe which normally executes a regular HTTP request, but is not available here yet. This ensures that we only put instances in rotation that have access to the configured model. As a result, a deploy with invalid environment will fill and leave the old instances in rotation.

To use this in development, you need to be signed in using gcloud to an account that has access to Vertex. At GitLab we use the ai-enablement-dev-69497ba7 GCP project for this.

For gitlab-org/gitlab#478883 (closed)

fix: patch ChatAnthropicVertex._format_output to include content

The original implementation for ChatAnthropicVertex._format_output could return a message that only has tool_calls populated but without content. This causes the following error message when calling Anthropic:

Error code: 400 - {'type': 'error', 'error': {'type': 'invalid_request_error', 'message': 'messages.16: `tool_result` block(s) provided when previous message does not contain any `tool_use` blocks'}}

To fix this, we've looked at the original implementation of ChatAnthropic._format_output, and patched the implementation we use here for switching to using Vertex.

We need to upstream a version of these change and upgrade our library. We'll do that in #96 (closed)