Support optional backend service for RAG-powered context retrieval

[scastlara]

Is your feature request related to a problem? Please describe.
lgtm-ai is currently a CLI tool. Its stateless, local nature makes it easy and fast to use, but also limits its functionality. Specifically:

• Advanced strategies for context retrieval (beyond the PR files) are not possible or hard to implement (see Improve code context given to the LLM #62).
• Real-time or automated webhook-driven interactivity (e.g., updating context on new commits, responding to comments [Add command to explain piece of code in a PR #48]) cannot be supported.
• Real time(-ish) feedback on reviews and comments (👍 / 👎 ).

Describe the solution you'd like
Introduce an optional backend service for lgtm-ai, while keeping the CLI as the frontend. The backend would provide:

1. Project onboarding – allow repositories to be registered with the backend.
2. Vector DB indexing – index repository content (code, docs, PRs/issues) for retrieval.
3. Webhook support – listen for events like commits, PR merges, comments, and update the index or trigger actions.
4. Context retrieval endpoint – given a PR diff, return the top-K relevant context chunks from the vector DB using RAG.

The CLI could then optionally query this backend when configured, while still working fully offline in local-only mode.

Possible lgtm config

[context]
use_backend_service = true
project_id = "xx"

use_backend_service = false (default) → CLI works as today (local, PR files only).
use_backend_service = true → CLI fetches enriched context from backend for RAG-powered reviews.

Describe alternatives you've considered
#62