feat: update Wafer provider models

[Jercik]

What changed

Updates the Wafer provider from the current https://pass.wafer.ai/v1/models response and follow-up live API checks.

• Adds deepseek-v4-flash, deepseek-v4-pro, and qwen3.7-max
• Updates pricing, cache-read pricing, context/input/output limits, and last_updated values
• Uses live /v1/chat/completions behavior for fields where /v1/models disagreed with runtime behavior

How I checked it

I fetched /v1/models with a Wafer Serverless API key, then tested the affected models through /v1/chat/completions.

The runtime checks covered text completion, tool calling, reasoning output, image input, video input, and large text prompts.

Findings that changed the model files:

• Qwen models returned reasoning_content even when /v1/models reported reasoning: false.
• Kimi and Qwen accepted image and video inputs. The video check used a small MP4 with a known code in the frame.
• DeepSeek V4 returned no reasoning by default, but did return reasoning_content when the request included reasoning_effort. Streaming responses exposed it as delta.reasoning_content.
• Qwen3.6-35B-A3B reports 256000 context in /v1/models. A full-context streaming probe with enable_thinking=false returned 190464 prompt tokens, 65536 completion tokens, and 256000 total tokens with finish_reason: "length".
• Default reasoning-mode large prompts for Qwen3.6-35B-A3B rejected with Range of input length should be [1, 229376], so the model file keeps context = 256_000, adds input = 229_376, and keeps output = 65_536.

Validation

• bun validate