Implement L4S support and the Prague CCA - #2729

TCP Prague emerged as the de facto implementation for scalable
congestion control defined by the Low Latency, Low Loss, and Scalable
Throughput (L4S) architecture (see RFC 9330
<https://datatracker.ietf.org/doc/html/rfc9330>).

The need for L4S is motivated by the desire to prevent the build-up of
large network queues that add significant delays to packets, while at
the same time avoiding underutilization or severe losses (due to
tail-drops or AQM-elicited early drops).

In essence, L4S builds on three pillars:

1) support from the network: network queues in routers and other nodes
   along a given network path capable of isolating L4S-capable flows and
   treating their ECN marking logic according to the specification

2) support from the protocols: reinterpretation of ECN bits in the
   network- and transport-layer headers

3) support from the hosts (endpoints): use of so-called scalable
   congestion control algorithms that react to congestion signals
   (primarily ECN-CE marks on packets) according to the behavior defined
   in L4S (e.g., scaling back the congestion window in fractional
   amounts proportionally to the ratio of ECN-CE marked packets seen
   recently -- which, in a properly functional network queue, is
   proportional to the current queue length compared to the queue's
   capacity)

Of course, Quinn is only concerned with the third point. Thus, this
change implements the Prague congestion control algorithm (CCA) in Quinn
to provide basic support for L4S, following the TCP Prague draft
<https://datatracker.ietf.org/doc/draft-briscoe-iccrg-prague-congestion-control/04/>.

Implementation notes:

- NewReno is used as Prague's inner CCA, since its RTT-independent
  multiplicative-decrease behavior is a closer match for what the
  scalable congestion control response in L4S expects.

- Congestion window reduction is paced on RTT, and general packet pacing
  for Prague-controlled flows is aligned with the draft's requirements.

- Loss detection and ECN processing for the same incoming ACK can both
  fire `on_congestion_event` for what is really a single queue-overflow
  event (`detect_lost_packets` runs ahead of `process_ecn` in
  `Connection::on_ack_received`, and many AQMs drop and CE-mark off the
  same queue-occupancy signal). Without mitigation this compounds to a
  ~75% window reduction instead of the ~50% floor required by the TCP
  Prague draft §2.4.1. The ECN cooldown is now armed from the loss event
  itself, mirroring the existing handling of the reverse (ECN-then-loss)
  ordering.

- Basic tracing for congestion window reduction and qlog emission of
  L4S-type events are added for observability.

Co-Authored-By: Felician Nemeth <[email protected]>

Add ECN configuration, Prague support to `quinn-perf`.

Changes:

- Expose the `--ecn` CLI option to set the ECN mode.

- Add `Prague` to the `--congestion` algorithm choices.

- Track and report RTT and RTT variance metrics per stream to analyze
  latency characteristics under L4S similarly to iperf3.

- Document L4S ECN requirements and configuration guidelines.

Also includes two unrelated fixes surfaced while working on the above:

- Close connections immediately once completed to prevent waiting on the
  30-second idle timeout, which caused perf runs to be needlessly slow.