feat(backend): add multi-provider LLM support via LLM_PROFILE

[dimknaf]

Summary

Adds five new LLM provider profiles alongside deepinfra, so the agent backend can run on any of:

Profile	Model	Required env
gemini	gemini/gemini-3.1-flash-lite-preview	GEMINI_API_KEY
deepinfra	deepinfra/google/gemma-4-31B-it	DEEPINFRA_API_KEY
nim	nvidia_nim/google/gemma-4-31b-it	NVIDIA_NIM_API_KEY
together	together_ai/google/gemma-4-31B-it	TOGETHER_API_KEY
local	openai/qwen (llama.cpp / vLLM / LM Studio on :8003)	LOCAL_API_KEY + LOCAL_API_BASE
local_gemma	openai/cyankiwi/gemma-4-31B-it-AWQ-4bit (Gemma 4 31B AWQ on :8002)	LOCAL_API_KEY + LOCAL_API_BASE_GEMMA

Switching providers is a single env-var flip (LLM_PROFILE=...) — no other config changes required.

What's in the diff

• backend/API/config.py — six profiles in _LLM_PROFILES; new agent_api_base setting + resolved_api_base property; resolved base URL logged on startup.
• backend/API/agent.py — _create_model() passes base_url=resolved_api_base to LitellmModel when the active profile defines an api_base_env. Cloud profiles behave exactly as before (base_url is omitted).
• docker-compose.local.yml — declares host.docker.internal:host-gateway on the backend service so containerised runs can reach a local LLM running on the host. Docker Desktop already maps this; the explicit entry gives Linux parity.
• backend/README.md + backend/.env.example — provider matrix and one-flip workflow documented.

Verification

• ✅ pytest — all 56 backend tests pass.
• ✅ LLM_PROFILE=deepinfra end-to-end via Docker Compose (cloud Gemma 4 31B).
• ✅ LLM_PROFILE=local_gemma end-to-end via Docker Compose (AWQ Gemma served by vLLM on :8002, reached through host.docker.internal).

Test plan

• Pull the branch, copy backend/.env.example to backend/.env, fill in DEEPINFRA_API_KEY, run docker-compose -f docker-compose.local.yml up --build. Hit /api/agent-status, expect "llm_profile": "deepinfra" and "litellm_model_initialized": true.
• Run a single CSV row from the UI; result populates as before — no regression for the default cloud path.
• (Optional) Stand up any OpenAI-compatible server on the host, set LLM_PROFILE=local (or local_gemma) with the matching LOCAL_API_BASE*, restart the backend with docker compose up -d (no rebuild), repeat — same result over the local endpoint.