Add a minimal web chat UI served by ds4-server

[phierru]

I added a minimal web chat interface, served directly by ds4-server at /, so a
model can be used from a browser without any extra tooling. There is no build step
and nothing to install: the page is embedded in the binary as a string literal, and
chat-ui.html is the same page as a standalone file if you prefer to open it from
disk.

Trying it

./ds4-server -m ./ds4flash.gguf --ctx 32768

Then open http://127.0.0.1:8000 — the UI is served by ds4-server itself on its
usual port, not by a separate process.

Two prompts that exercise the rendering:

1. Show me the Maxwell equations
2. Explain briefly how these equations imply electromagnetic waves

The first comes back with headings, bold text and display equations; the follow-up
uses inline math inside bullet lists.

What it does

• Streams responses over the existing /v1/chat/completions SSE endpoint.
• Renders Markdown: headings, nested lists, tables, blockquotes, links, horizontal
  rules, fenced and inline code.
• Renders LaTeX with KaTeX: $$..$$ and \[..\] as display math, $..$ and
  \(..\) inline. Inline $..$ requires a non-space character next to both
  delimiters, so ordinary prose like "costs $5 and $7" is not swallowed as math.
• A thinking selector in the header sends reasoning_effort per request, off by
  default. GLM 5.2 emits visible reasoning otherwise, which is tedious to read at
  streaming speeds.
• A copy button per message, which copies the raw Markdown rather than the rendered
  HTML, and a stop button (also bound to Esc) that aborts generation while keeping
  whatever text already arrived.

Notes, and things you may want changed

• KaTeX is loaded from a CDN (jsDelivr). It cannot reasonably be inlined because
  of the fonts. If it fails to load, equations degrade to showing the raw LaTeX
  instead of breaking, and re-render if it arrives late. I am happy to drop math
  rendering altogether if you would rather the UI carry no external dependency.
• The copy embedded in ds4_server.c is generated from chat-ui.html rather than
  escaped by hand, since the LaTeX regexes are backslash-heavy and one missed escape
  yields a page that compiles but renders wrongly. tools/gen_chat_ui.py regenerates
  it; tools/gen_chat_ui.py --check reports staleness without writing, if you want
  it in a pre-release check.
• tools/test_chat_ui.js covers the renderer (Markdown blocks, HTML escaping, math
  delimiters). It takes an optional path, so the same suite can be run against a page
  fetched from a running ds4-server — that is what verifies the C escaping survived
  the round trip, which compiling cannot tell you. Plain node, no dependencies.
• Tested on macOS / Metal against DeepSeek V4 Flash (resident) and GLM 5.2 (SSD
  streaming). The UI is model-agnostic and reads the model id from /v1/models.
• Purely additive: no existing file behaviour changes, and the new route only
  answers GET /, which was previously unhandled.