fix(http): add connection-level timeout protection against non-reading clients

[bug-ops]

Summary

• apply_common_layers' TimeoutLayer/ResponseBodyTimeoutLayer (fix(http): remove dead frame streams, add timeout/concurrency limits #521) bound handler execution and the pre-response phase, but a client that stops reading the response socket is never caught: ResponseBodyTimeoutLayer's deadline only resets on body poll, and hyper stops polling once its write buffer fills waiting on the client to read. Closing that gap requires accounting at the TcpListener accept loop, which pjs-core previously had no API for (it only builds Routers; axum::serve is called by the embedding process).
• Adds infrastructure::http::serve::{ConnectionLimits, serve_with_limits}, an http-server-gated replacement for axum::serve built on hyper_util::server::conn::auto::Builder. It bounds per-connection lifetime with a deadline around the connection future (default 300s, independent of body-poll activity — this is what closes the HTTP server has no connection-level protection against a non-reading client (residual #515 gap) #523 threat scenario), bounds header read time via header_read_timeout (default 10s, ~6x tighter than hyper's own 30s default, fixing a separate slowloris gap discovered during this work since axum::serve never configures a hyper timer), and caps concurrent connections via a semaphore acquired before accept() (default 1024, a distinct axis from fix(http): remove dead frame streams, add timeout/concurrency limits #521's handler-concurrency cap).
• ConnectInfo<SocketAddr> is preserved per connection so the crate's own WebSocket upgrade handler and per-IP rate limiter keep working with the new serve function.
• Accept-loop errors now retry with backoff (mirroring axum::serve's own behavior) instead of killing the whole server on a single transient error.
• Wires all three pjs-demo server binaries to serve_with_limits in place of axum::serve; websocket_streaming.rs sets max_connection_duration: None to protect its long-lived /ws/{id} route.

This went through a full security-audit → implement → adversarial critique → fix → re-verify → code-review cycle. The critique caught two real correctness gaps in the first pass (dropped ConnectInfo, accept-error propagation killing the server) — both fixed and independently re-verified before this PR.

Test plan

• cargo +nightly fmt --check
• cargo clippy --workspace --all-targets --all-features -- -D warnings
• cargo nextest run --workspace --all-features --lib --bins
• cargo test --workspace --doc --all-features
• RUSTFLAGS="-D warnings" RUSTDOCFLAGS="--deny rustdoc::broken_intra_doc_links" cargo doc --no-deps --workspace --all-features
• 6 new integration tests in crates/pjs-core/tests/http_serve_connection_limits.rs, including a regression test for the literal HTTP server has no connection-level protection against a non-reading client (residual #515 gap) #523 threat scenario (client completes a request against a large-body route, then stops reading — connection is force-closed within the configured deadline) and one for ConnectInfo availability to handlers
• All three pjs-demo binaries build and run under the CI feature set

Closes #523
Closes #515

[github-actions]

WASM Bundle Size Report

Target	Raw (KB)	Gzipped (KB)	Status
web	154	69	PASS
nodejs	154	69	PASS
bundler	154	69	PASS