cmd/tailcat, internal/perf: add iperf-like perf command and service

[bradfitz]

Add "tailcat perf " and a matching "perf" service for
"tailcat serve" to measure what a tunnel can carry: TCP or UDP
throughput in either or both directions, with one or more parallel
streams, for a duration or a byte count. UDP tests report loss,
reordering, and RFC 3550 jitter from a sequence number and send
timestamp in each datagram. All tests report round-trip latency
measured over the test's control connection while the data streams
are running, which is the number that matters when a full pipe makes
everything else sluggish.

The protocol lives in internal/perf. The client opens a JSON-lines
control connection to TCP port 5201 on the server's tailcat address,
data flows on separate TCP connections or UDP flows to the same port,
and each side reports what it sent and received so the client prints
both views of every direction. The server runs one test at a time
and bounds stream count and duration. The service plugs into the
existing OnTCP and OnUDP hooks with the packet filter narrowed to
that port, and logs one line per completed test.

A throughput test through a shared DERP relay mostly measures the
relay's rate limit while crowding out everyone else on it, so the
client waits for a direct path first (up to --timeout, like
ping --until-direct) and refuses to run relayed. --via-derp allows a
relayed test through a relay the user runs, but never through one
under ipn.dev or tailscale.com. Client.DERPRegion is added to the
library so the CLI can see which relay is in use.