Add Gonka provider for Kimi K2.6

[knyazev741]

What

• add a dedicated Gonka provider exposing only moonshotai/Kimi-K2.6
• keep the broker model ID, base URL, API key, weight, context length, and blended token cost operator-controlled
• stream OpenAI-compatible responses, assemble tool-call chunks, require usage metadata, and account for cost
• register Gonka in the inference gateway without advertising embedding support or adding a fallback
• add Gonka to local miner provider selection and reject embeddings before any network call
• document safe disabled defaults in both gateway and miner env examples

Why

Gonka-compatible brokers can expose Kimi K2.6 at a substantially different cost profile from existing providers. A first-class provider keeps credentials and routing under gateway control while preserving Ridges' model whitelist and evaluation cost accounting.

Operator configuration

• USE_GONKA
• GONKA_BASE_URL
• GONKA_API_KEY
• GONKA_WEIGHT
• GONKA_KIMI_MODEL_ID (defaults to moonshotai/kimi-k2.6)
• GONKA_CONTEXT_LENGTH (defaults to 240000)
• GONKA_COST_USD_PER_MILLION_TOKENS

Validation

• python -m pytest -q — 512 passed
• focused provider/miner tests — 20 passed
• Ruff checks for the new provider/tests and import ordering across modified Python files
• formatter and git diff --check

No live broker credentials or paid inference calls were used.