The 10-second read deadline set for SNI peeking in the SNI-deferred
path was cleared via defer, which only runs when handleConnect returns.
Since handleConnect blocks in relayData for the connection lifetime,
the deadline persisted and killed every SNI-deferred connection after
10 seconds. This caused streaming API responses to be truncated
(manifesting as OpenAI "terminated" errors), tool call fetches to fail,
and periodic TLS handshake failures on chatgpt.com.

Fix: clear the deadline explicitly after SNI peek completes, before
the relay phase begins.

Also fixes two secondary issues:

- relayData socket leak: when the first relay direction completed,
  the function could block indefinitely waiting for the second
  goroutine if goproxy held the MITM connection open. Close writer
  and set a read deadline on target to force cleanup. Eliminates
  CLOSE_WAIT socket accumulation (75 leaked sockets observed).

- goproxy Transport stale connections: the MITM proxy Transport had
  no IdleConnTimeout, causing dead pooled connections to persist
  indefinitely. Add IdleConnTimeout (90s) and MaxIdleConnsPerHost (4).

- goproxy log noise: suppress expected broken pipe and handshake EOF
  warnings via a filtered logger. These are normal for short-lived
  polling connections (Telegram getUpdates).