feat: llm e2e tests

[gregnr]

Adds a new test suite for end-to-end LLM-based tests. Simulates real world MCP usage by:

1. mocking orgs / projects / databases via mock management API (as we already do for unit tests)
2. using ai-sdk to orchestrate real text generation via an LLM
3. connecting the MCP server via ai-sdk's MCP client
4. testing typical chat questions / prompts users use when interacting with Supabase MCP
5. evaluating which tools were called as a result of the user prompts
6. evaluating the text response using llm-as-a-judge concepts

To facilitate llm-as-a-judge, we create a custom vitest extension called toMatchCriteria() that accepts a natural language criteria string and asks an LLM to evaluate the generated output against the expected criteria. Eg:

await expect(text).toMatchCriteria(
  'Describes a single table in the "todos-app" project called "todos"'
);