docs(qwen3): add 14B and GH200 family benchmarks

[FeathBow]

Description

Expands the Qwen3 dense-family page to cover 4B, 8B, 14B, and 32B, and restructures Performance into two clearly separated suites.

• Adds Qwen3-14B: a launch subsection (the 40-query-head / 8-KV-head GQA group has no compiled decode kernel, so decode reroutes through the batched eager prefill path, logged as a WARN at load), a serving ladder, architecture notes, the sidebar label, and the overview table row.
• Separates the historical RTX 5090 openinfer-vs-vLLM comparison from the new GH200 results. The 5090 tables keep their numbers and gain a harness note: one seed for every sweep point against one long-lived server with the prefix cache on, so both engines saw the identical prompt stream and the side-by-side is like-for-like, but absolute numbers are order-dependent and not directly comparable to re-runs under the current per-point-seed script.
• Adds a GH200 cold family-scaling section: one measurement contract (openinfer main c116077b, fresh server per sweep point, seed 42, Python vllm bench serve, random 1024/128, greedy), a summary table (decode path, high-load throughput, queueing knee), then full ladders for 4B / 8B / 14B / 32B. The 4B/8B ladders are family-scaling anchors, not replacements for the RTX 5090 comparison. 14B and 32B keep their earlier-run evidence (long-context, overload, tool-call, greedy-vs-HF, logits golden gate) with explicit attribution.
• Styling: adds an h4 rule and restores full-width table layout on wide screens (Starlight defaults tables to display: block, which kept width: 100% from stretching the cell grid; narrow screens keep the scrollable block), and raises this page's TOC depth to h4.
• Validation: npm run build; static tests 12/12; table layout and TOC checked in a browser at wide and narrow viewports.