Run the HTTP/3 tests against a local server, fix two HTTP/3 bugs

[benoitc]

Nine test modules sent HTTP/3 requests to public servers (mostly cloudflare.com) under eunit's default 5 s limit, and accepted a failed connection as a pass. A slow UDP path from a CI runner made them time out: that is the macOS flake on #939 and #941.

hackney_h3_test_server starts an in-process quic_h3 server on 127.0.0.1 with the test certificates. The tests now assert real results instead of tolerating network errors: protocol, status, headers and bodies, redirect following and the max_redirect limit, Alt-Svc upgrading a TCP origin to HTTP/3 and the blocked-host fallback, and certificate verification against the test CA. The public-server checks in hackney_http3_e2e_SUITE move to an interop group that runs with HACKNEY_H3_INTEROP=1.

Three HTTP/3 bugs the stricter tests turned up, fixed here:

• The response to an HTTP/3 streaming upload could not be read: body/1 returned {error, invalid_state} and stream_body/1 returned {error, no_stream}, because the body went to the start_response/1 caller as a second reply and the stream was dropped. The stream now enters the same pull states as request_streaming, body/1 gains an HTTP/3 clause, and start_response/1 answers at once when the headers came first. finish_send_body/1 no longer parks its already answered caller, which made start_response/1 hang when the headers arrived early.
• hackney_h3 passed cacertfile to quic, which only takes DER cacerts, so it was ignored and verification failed as {error, timeout}. It is now decoded.
• A closing HTTP/3 connection reported no reason. quic_h3 sends one with its close event and hackney only matched the older shape without it, so the message was dropped, the h3 process stayed alive and the caller waited out its timeout. A handshake that fails on a bad certificate now comes back as {error, {connection_closed, {certificate_invalid, _}}} instead of {error, timeout}. Both shapes are handled, since quic =< 1.10.0 still closes without a reason on two of its paths.

Found along the way, left for separate changes:

• Resuming an HTTP/3 connection from a cached session ticket can stall the handshake: 6 of 10 runs on one scheduler, none with zero_rtt off. The local tests run with {zero_rtt, false}.
• When the HTTP/3 connect times out, hackney:connect lets the gen_statem:call timeout exit escape instead of returning {error, _}.
• Once a quic release carries Finish the HTTP/3 owner contract and accept a CA file erlang_quic#355, hackney should answer callers waiting in stream_body/1 or body/1 when their stream is reset. Until then quic drops a peer RESET_STREAM on request streams, so that path cannot be tested.