docs(qwen3-32b): add page with single-GH200 serving numbers

[FeathBow]

Summary

• Add Qwen3-32B page — launch commands, serving numbers on a single GH200 120GB (QPS sweep, saturation ~5.3 req/s / ~680 out tok/s), architecture notes. Measured on openinfer main 5959f05.
• Supported-models row — index table links the new page.
• Sidebar entry — Models group now lists Qwen3-32B.

Test plan

• Rendered page reviewed at localhost (npm run build passes, 8 pages)
• All commands on the page are runnable (download, launch, curl)
• Benchmark data matches the run artifacts posted in the verification report

[xiaguan]

Thanks for the contribution! The benchmark methodology is solid and the HF correctness comparison is a nice touch. One structural change before we merge:

Merge into the Qwen3 family page

The existing qwen3-4b.md already serves as the Qwen3 family page — it hosts both 4B and 8B under one file, with 8B as a ### Qwen3-8B subsection (see qwen3-4b.md:65). Since your page explicitly notes 32B is "the same qwen3 crate ... same model_type: qwen3 ... no feature flags or build changes needed," it fits the same pattern.

Please merge the 32B content into qwen3-4b.md as a ### Qwen3-32B subsection (same shape as the 8B subsection), and delete src/content/docs/models/qwen3-32b.md. This keeps "one Qwen3 family = one page" consistent so future contributors know where Qwen3-N-B belongs.

Concretely:

• Move the GH200 benchmark table, launch command, tool calling example, and architecture notes under ### Qwen3-32B in qwen3-4b.md.
• Remove the standalone qwen3-32b.md file.
• Drop the sidebar entry added in astro.config.mjs (the existing Qwen3-4B entry covers the family).
• In index.mdx, keep the supported-models table as a single Qwen3 row pointing at /models/qwen3-4b/.

This also means tests/routes.mjs doesn't need changes (no new route), which avoids the silent-CI-gap that would've existed with a standalone page.

Minor

index.mdx architecture column mixes axes. The current Qwen3-4B/8B row says "Full attention, tensor parallel" (a capability). If you keep that phrasing for the merged Qwen3 row, note that 32B also supports TP (same model_type). Alternatively, if the column describes the benchmark setup, both rows are actually TP1 single-GPU — pick one axis so readers aren't misled into thinking 4B/8B rely on TP while 32B relies on large VRAM.

[FeathBow]

Thanks for the look :) Tool calling seems a general capability shared by all Qwen3 sizes, should it be pulled out into its own section rather than sitting under Qwen3-32B?