Parse the request body before raising HttpParserUpgrade

[Kludex]

Fixes #124.

cb_on_headers_complete returns 1 whenever the upgrade flag is set. That tells llhttp to skip the body, so an upgrade request that has a body (e.g. Upgrade: h2c + Content-Length or chunked, which JDK's HttpClient and curl --http2 send by default) raises HttpParserUpgrade right after the headers. The body is lost, and the bytes after the offset are parsed as a new request.

llhttp already makes this decision itself in llhttp__after_headers_complete: it pauses right after the headers for upgrades without a body, CONNECT, and 101 responses, and otherwise parses the body and pauses after on_message_complete. Returning 0 lets it do that.

What changes:

• An upgrade request with a body fires on_body and on_message_complete first. Then HttpParserUpgrade is raised with the offset after the body.
• WebSocket handshakes (GET with no body), CONNECT, and 101 responses behave the same as before.

This matches the RFCs. A server may ignore Upgrade (RFC 9110 §7.8), and an h2c upgrade request's body has to be read as HTTP/1.1 before switching protocols (RFC 7540 §3.2).

I ran uvicorn's protocol test suite against this branch and it passes. It also fixes Kludex/uvicorn#2722.

AI Disclaimer

This PR was developed with the assistance of either Claude or Codex. I've reviewed and verified the changes.