feat: record the first join as a latency DAG (3RTT T01)

[tbarbugli]

3RTT task T01. Replaces ConnectionTiming (folds in #2) with jointrace: the first join is recorded as a DAG of spans, so a trace shows which steps run in parallel, which chain is critical, and every network step in round trips of its peer. Builds on #5.

What's in it

• jointrace package: Span{Name, After, Parent, Start, End, Kind, Peer, Note}, with the node names from research/join-dag.md as constants.
  • Trace.CriticalPath() walks back from the terminal step (pub.rtp/sub.rtp), following the dependency that finished last.
  • Trace.String() draws the DAG with a gantt column and marks the critical path; json.Marshal(trace) gives a flat Report.
• Call.JoinTrace() and Call.OnJoinTrace(fn). The handler fires once media flows both ways, or after JoinTraceTimeout. Reconnects after the first join seal the trace.
• httptrace detail (.dns, .tcp, .tls, .request, .first_byte) on:
  • the coordinator REST join,
  • the coordinator and SFU websockets, via the new internal/wsdial,
  • the SFU twirp RPCs.
• A .server span from Server-Timing when the server sends one.
• RTT per peer: coordinator and SFU from the TCP connect, UDP from the ICE selected pair.
• WithNetworkDelay(rtt), for tests and benches only. Every connection the client opens is delayed rtt/2 each way, in order, and a TCP connect takes one RTT. This covers coordinator REST and WS, the location hint, SFU WS and twirp, and ICE UDP via SettingEngine.SetNet.
• signal.WithDialContext, signal.WithTLSConfig, testutil.WithTLS().

Tests

• jointrace: the critical path through parallel branches, gaps counted as wait, a JSON round trip, a render golden (go test ./jointrace -update), httptrace phases and Server-Timing over a delayed TLS server, a traced wss dial at 3 RTT, and a recording cost check (< 1 ms per join).
• internal/netdelay: TCP connect and echo, and UDP echo, each 100 ms ± 5%; order kept over 500 writes; deadlines.
• A full join against fakesfu with real pion peers both ways:
  • Every applicable node is present, each after its dependencies.
  • The critical path total is within 5 ms of Join → last first RTP.
• The same join with WithNetworkDelay(100ms) over TLS: sfu.ws.dial is 3 RTT and sfu.join is 1 RTT, each ± 10%, and the SFU and UDP RTTs are 100 ms ± 10%.
• make ci is green.

Example (fake SFU over TLS, WithNetworkDelay(100ms)):

critical path: pcs.create > sfu.ws.dial > sfu.join > sub.debounce > sub.offer > sub.sendanswer > sub.ice > sub.dtls > sub.rtp
  1405.4 ms = 10.83 RTT (1089.4 ms network) + 314.3 ms timers + 1.8 ms local; 0.0 ms unaccounted waiting