Fix Telegram aiogram session compatibility - #13

Summary

• align _TelegramPollingSession with current aiogram session interfaces
• remove the stale startup watchdog hook that still referenced deleted polling methods
• add focused Telegram regression coverage for the session wrapper and startup path

Problem

On hosts with transports=["feishu", "telegram"], Telegram transport startup could crash before the service came online after aiogram interface changes. In live impact, this caused ControlMesh user services to loop under systemd instead of serving the primary transport.

Changes

• implement _TelegramPollingSession.stream_content() by delegating to the wrapped aiogram session
• update _TelegramPollingSession.make_request() to accept and forward timeout= instead of the old request_timeout= positional path
• stop run_startup() from scheduling the removed _watch_polling_liveness() / _polling_watchdog path
• add focused tests for timeout forwarding, stream delegation, and startup task wiring

Validation

• user requested immediate PR; I did not block on a full suite
• focused local pytest slices were partially run during development, but this PR is intentionally not gated on finishing broader test execution