feat(blueclaw): add Blue Claw provider

[eliteprox]

Adds BlueClaw, an OpenAI-compatible provider built on the Livepeer network (https://openai.blueclaw.network/v1) serving open Qwen models. Auth is a bearer token from BLUECLAW_API_KEY.

Two models, both verified live against the endpoint:

• Qwen/Qwen3.6-35B-A3B-FP8 — context 131072 (read from the vLLM max_model_len error)
• Qwen3.6-27B — context 196608

For each I confirmed chat completions, streaming, and tool calls (finish_reason: tool_calls with a real function call). Reasoning comes back in a separate reasoning field; interleaved is omitted since that field name is not one of the two the schema allows. No cost block — free fair-use beta with no published pricing.

All four review items from #2777 are addressed:

• Logo: added providers/blueclaw/logo.svg (PNG-embedded SVG from the site favicon — no vector asset is published; same approach as orcarouter/atomic-chat/hpc-ai)
• base_model: both models now inherit from alibaba/qwen3.6-35b-a3b / alibaba/qwen3.6-27b, overriding only what the endpoint actually differs on (text-only input, attachment = false, verified narrower context)
• reasoning_options: explicit [[reasoning_options]] type = "toggle" on both, matching how evroc/nearai declare the same FP8 model
• limit.output: no longer copied from the combined context length; inherits the base model's 65_536

[rekram1-node]

Automated review:

Two issues need addressing before merge:

1. providers/blueclaw/logo.svg embeds a colored PNG and hardcodes width/height. New provider logos must be vector SVGs without fixed sizing or hardcoded colors, using currentColor so they adapt to light and dark themes. Please replace this with a compliant SVG.

2. Both models declare reasoning_options = [{ type = "toggle" }], but the PR only cites equivalent Qwen deployments on other providers and observation of a separate reasoning response field. Reasoning controls are provider-specific. Please provide Blue Claw request-level evidence showing the exact API syntax that enables/disables reasoning, or use reasoning_options = [] when no control is verified.

The base_model usage and focused provider configuration otherwise look good.