Add LM Studio as a keyless local model provider.

[sobiya-22]

Fixes #122

Summary

Adds LM Studio as a first-class keyless local model provider, following the same pattern as Ollama.

• Register lm_studio in the provider registry with default endpoint http://localhost:1234/v1
• Wrap OpenAIProvider with placeholder key lm-studio and configurable base_url
• Keyless verification via GET /v1/models (no auth header)
• Live model discovery and liveness gating for lm_studio:* entries in the model picker
• Conservative capability defaults for local models (same as Ollama)
• Provider-specific Settings help text and gallery ordering

LM Studio models are not added to matrix.py — users load arbitrary GGUF models locally. Models can be discovered after Detect (when the local server is running) or added manually via the existing "Add another model…" field.

Test plan

• pytest tests/test_provider_router.py tests/test_provider_verify.py tests/test_settings.py tests/test_server.py
• Start LM Studio, load a model, enable local server
• Settings → LM Studio → Detect → card shows "✓ Running"
• Configure Models → suggested loaded models appear → tick one → use in composer as lm_studio:<model>
• Send a chat message and confirm completion works

[LifetimeVip]

Support

[zalex666]

Nice work on this — and apologies, I opened #434 for the same feature before spotting this PR (should have searched first). Laying out the delta here so maintainers can pick a direction easily.

#434 follows the same core pattern (keyless descriptor + OpenAIProvider at <root>/v1, liveness-gated picker entries via /v1/models) and additionally covers:

• Keyless first-Detect persistence. With the endpoint field untouched, a passing Detect stores no profile, and _lm_studio_models() gates on a stored profile — so the live model list and suggestions stay empty until the endpoint is manually edited. (Add LM Studio as a local model provider (and fix keyless-provider discovery) #206 described the same latent bug for Ollama.) Fixed GUI-side (Detect always persists for keyless providers) plus backend (is None gate so an empty stored profile still probes the default URL) — that's what makes "Detect → suggested models appear" work at the default URL.
• model_ready and the default-model handoff gate on actual liveness (_provider_available), so a dead local server no longer reads as ready and no longer blocks a working provider from taking the default.
• Probe/verify check the response shape (data/models list), not just any 2xx — a non-LM-Studio service on :1234 no longer passes Detect.
• Liveness-cache invalidation on endpoint changes; an explicitly-cleared endpoint verifies the default instead of the stored URL; failed saves surface in the GUI instead of showing "✓ Tested & saved".
• The lobe-icons brand mark (this branch adds lm_studio to PROVIDER_ORDER but not PROVIDER_LOGOS, so the card renders the monogram fallback), a README mention, and a recommended_model for the post-Detect auto-add.
• Rebased on current main (this branch is showing conflicts at the moment).

Validated #434 live against LM Studio: tool-calling round trips on two models incl. JIT load and a thinking model's reasoning_content streaming.

Zero attachment to which PR lands — if maintainers prefer this one, everything above can be cherry-picked here and I'll close #434.