feat: hybrid LLM model picker + per-model cost tracking

[AlanY1an]

Summary

• Replace the admin LLM card's free-text model input with a hybrid picker (curated dropdown for official Anthropic / OpenAI providers + Custom… escape hatch + free-text fallback for non-official base_url). base_url and api_key_env are also editable inline.
• Migrate the cost-tracking ledger from role-only rates to per-(provider, model) pricing via a vendored copy of LiteLLM's model_prices_and_context_window.json (MIT-licensed; manual refresh via scripts/refresh_litellm_prices.py). Stub provider and local-host base_urls short-circuit to $0.
• PATCH-time validation of llm.api_key_env returns HTTP 422 with structured [{field, msg}] so the admin UI renders the error inline. Skipped for provider="stub" and local-host base_urls.

Test plan

• `uv run pytest` (expect 1444 passed)
• `uv run ruff check src/ tests/`
• `uv run lint-imports`
• `cd src/echovessel/channels/web/frontend && npx tsc -p tsconfig.json --noEmit`
• `uv build` and confirm the wheel bundles `echovessel/core/llm/data/{litellm_prices.json,LITELLM-LICENSE}`
• Browser: provider=anthropic, empty base_url → picker shows Sonnet / Haiku / Opus with prices inline
• Browser: pick `Custom…`, type a non-preset model id, save, reload → value persists, dropdown shows Custom selected
• Browser: set base_url to `https://openrouter.ai/api/v1\` → picker collapses to free-text with the "custom base_url detected" hint
• Browser: set base_url to `http://localhost:11434/v1\` with a missing api_key_env → save returns 200 (local-host skip)
• Browser: change api_key_env to a name not in environment → save returns 422 with inline error near the field; saving with a valid name clears it
• Browser: temperature / max_tokens / timeout / save-button enable-on-dirty all unchanged