feat: add optional remote compute backend

[Luis Serrano (luisnomad)]

Summary

Adds an optional remote-compute mode so LTX Desktop can keep its Electron UI and project files on one computer while running models and inference on a user-owned headless GPU machine.

• Keeps the existing managed-local backend as the default.
• Adds a first-run local/remote compute choice and a persistent Settings → Compute control.
• Adds an authenticated standalone FastAPI deployment mode for a headless GPU host.
• Transfers inputs and generated artifacts through authenticated media IDs instead of sharing client filesystem paths.

This implements the self-hosted split-UI/GPU-server scenario described in #82 (and related discussion in #53). That request identified authentication, transport security, packaging, and UI flow as the missing pieces; this PR covers each of those areas.

Design

Desktop client

• Stores the optional remote URL and bearer token through Electron; the token is encrypted with safeStorage.
• Tests protocol/capability compatibility before saving a remote connection.
• Routes backend HTTP, WebSocket, media upload, and artifact download through the active connection.
• Imports downloaded outputs into the existing local project-assets directory.
• Clearly separates the remote-backend access token from the LTX Cloud API key.
• Supports authenticated direct HTTP on a deliberately configured trusted LAN, with an in-app warning that the token and media are not encrypted in transit.

Headless backend

• Adds LTX_DEPLOYMENT_MODE=standalone with a required LTX_AUTH_TOKEN of at least 32 characters.
• Requires bearer authentication for HTTP and WebSocket requests; Basic auth is rejected in standalone mode.
• Disables the shutdown endpoint in standalone mode.
• Remains loopback-only by default. A non-loopback listener requires authentication and explicit LTX_BIND_HOST configuration.
• Documents direct trusted-LAN HTTP as the minimal private-network setup, SSH as an encrypted private-host alternative, and HTTPS for shared, untrusted, or internet-facing networks.
• Uses bounded uploads, validated media types, opaque artifact IDs, expiry cleanup, and path-containment checks.
• Returns an explicit protocol version and capability set for client compatibility checks.
• Keeps model paths and Hugging Face authentication owned by the compute backend.

See docs/REMOTE_BACKEND.md for deployment and connection instructions.

Compatibility

• Existing installations continue using the Electron-managed local Python backend unless remote compute is explicitly selected.
• LTX Cloud generation remains independent of this feature.
• Remote compute can be changed after onboarding through the status control or Settings → Compute.
• The backend remains a trusted, single-user service; this does not attempt to introduce multi-tenancy.

Validation

• pnpm typecheck
• pnpm backend:test — 285 passed
• pnpm build:frontend
• Packaged Apple Silicon client connected through an SSH tunnel to an Ubuntu/NVIDIA headless backend.
• Completed an image-to-video generation on the remote LTX model, downloaded the authenticated artifact, and imported it into the Mac project assets.
• Restarted the client with the saved remote configuration and confirmed that it reconnects without starting the Mac Python backend.
• Bound the backend to a specific trusted-LAN address and connected the Mac directly after stopping the SSH tunnel; an unauthenticated probe returned 401 and the saved authenticated desktop connection succeeded.

Notes for reviewers

The generated OpenAPI TypeScript/JSON changes account for a substantial portion of the diff. The implementation follows the existing route → handler → service structure, with a fake media service used by integration tests rather than mocks.