yandex: reduce reconnect churn (participant pile-up) and close stale conns

[D13tre33]

Summary

Follow-up to #54 (which fixes the yandex transport's handshake race).
With that race fixed, sessions establish reliably, but I noticed
something else while watching a live exit node for a while afterward:
a session that fully succeeds still eventually gets an explicit
server-side disconnect (disconnectReason code 4007, "drop") after
roughly 60-90 seconds.

While digging into that (not fixed here - see below), I logged the
server's auth/connectState messages, which list the doc room's
current participants. Every WebSocket reconnect - even ones that fail
fast - registers as a brand new "Anonymous" participant server-side
(Yandex assigns its own participant id per connection and ignores our
client-side user.id). During a failure streak I saw 2-3 simultaneous
ghost participants pile up in the same room from nothing but our own
reconnect attempts.

This PR doesn't fix the underlying 4007/drop behavior (that looks
like it's coming from Yandex's own collab-session participant limits,
not something client-side code controls), but it does reduce how fast
our own retries can pile up participants:

• reconnectBackoff()'s floor goes from 500ms to 1.5s and its cap
  from 15s to 30s. Successful connects never hit backoff at all, so
  this only affects failure streaks.
• The main read loop's error branch now calls conn.Close() instead
  of leaving the socket for the runtime to clean up whenever.

Test plan

• CGO_ENABLED=0 go build succeeds (via the repo's Dockerfile)
• Deployed as the exit-node profile; confirmed in logs that the
  new backoff floor/cap take effect on reconnect