feat: add Avian as LLM provider

[avianion]

Summary

• Add Avian as a new OpenAI-compatible LLM provider (https://api.avian.io/v1)
• New provider module (web/src/llm-api/avian.ts) following the same pattern as Fireworks/SiliconFlow, with streaming + non-streaming support, per-model pricing, usage/billing tracking, TTFT measurement, and error handling
• Route avian/* model IDs through the Avian provider in the chat completions API
• Add AVIAN_API_KEY to server env schema (optional, like other third-party provider keys)
• Register Avian models in model-config.ts constants, providerModelNames, providerDomains, and all three copies of the ModelName type

Models available via Avian

Model	Context	Input	Output
avian/deepseek-v3.2	164K	$0.14/M	$0.28/M
avian/kimi-k2.5	128K	$0.14/M	$0.28/M
avian/glm-5	128K	$0.25/M	$0.50/M
avian/minimax-m2.5	1M	$0.15/M	$0.30/M

Test plan

• Verify isAvianModel() correctly matches only avian/* model IDs
• Test streaming and non-streaming requests route to Avian when model starts with avian/
• Confirm billing/credit consumption works correctly with Avian pricing
• Verify AvianError is caught and returned with proper status codes
• Check that non-Avian models still route to their existing providers (no regression)

[hiSandog]

One thing I'd double-check before merging: web/src/llm-api/avian.ts currently bills as soon as any streamed chunk contains usage. In this repo, the newer provider implementations in web/src/llm-api/canopywave.ts and web/src/llm-api/siliconflow.ts added a billedAlready + isFinalChunk(...) guard because some providers emit cumulative usage on non-final chunks or can repeat a terminal usage frame.

Without that guard here, Avian could insert duplicate BigQuery rows and consume credits more than once for the same response. I'd be inclined to copy the CanopyWave/SiliconFlow pattern and strip usage from non-final chunks before forwarding them to the SDK.

[avianion]

Addressed feedback: added billedAlready + isFinalChunk() guard to prevent duplicate BigQuery rows and credit charges. Follows the same pattern used in canopywave.ts and siliconflow.ts — usage is now stripped from non-final chunks and billing only triggers once on the final chunk with finish_reason.