Problem

During network disruption, agent.Collect() spawns one goroutine per
active peer to call Ping(). All N goroutines attempt to acquire
magicsock.Conn.mu simultaneously. The 5-second pingCtx timeout is
ineffective because pingWithType spawns a fire-and-forget goroutine
that remains blocked on the lock after the context expires, still
tracked by the wgengine watchdog. With 4 peers, this creates ~80s of
blocked goroutines, exceeding the 45-second watchdog threshold and
killing the agent via log.Fatalf.

Solution

Add a capacity-1 channel semaphore so at most one Ping is in-flight at
a time. Each goroutine acquires the semaphore before calling Ping (with
a pingCtx.Done() fallback) and releases it after. This reduces
watchdog exposure from N * timeout to 1 * timeout.

The peer-ping loop is extracted into collectPeerLatencies() behind a
peerPinger interface to enable direct testing of the concurrency
constraint.

Testing

• New TestCollectPeerLatencies_SerializedPings: creates 5 fake peers
  with a 10ms ping delay, asserts max concurrent pings never exceeds 1
  using an atomic high-water mark counter.
• Verified the test fails without the semaphore (concurrent count = 5).
• Existing TestStatsReporter and TestAgent_Stats_* tests pass.

Refs #22864