fix(telegram): classify PTB heartbeat transport errors

[kshitijk4poor]

Summary

Telegram polling heartbeat failures wrapped by python-telegram-bot now enter the existing reconnect ladder, while semantic Bot API errors remain non-retryable.

Based on #62060 by @liuhao1024. The substantive production commit preserves the contributor's authorship.

What changed

• Route heartbeat errors from both get_me() and the pending-update probe through the shared _looks_like_network_error() classifier.
• Correct that classifier for PTB 22.6's hierarchy, where BadRequest inherits from NetworkError despite being a permanent request error.
• Reuse the classifier for Telegram initialization retries and the polling callback, so all reconnect decisions share one policy.
• Route heartbeat recovery through _schedule_polling_recovery() for existing-task dedupe and background-task lifecycle tracking.
• Make gateway Telegram mocks reproduce PTB's real exception inheritance.
• Add load-bearing behavior tests for transport and semantic errors across both heartbeat probe paths, plus shared-consumer guards.

Divergence from #62060

The original inline isinstance(NetworkError) classifier was not retained because it also matched PTB BadRequest. Its negative test used an unrealistic mock hierarchy and passed on unmodified main. This salvage fixes the shared classifier instead of adding a second policy.

Verification

• 98 passed across Telegram reconnect, pending-update, and thread-fallback suites.
• Real PTB 22.6 matrix on both heartbeat paths:
  • NetworkError, TimedOut -> one recovery
  • BadRequest, InvalidToken, Forbidden, RetryAfter -> no recovery
• Mutation check: removing semantic exclusions fails 3 regression tests.
• Ruff, compile, and diff checks pass.
• Codex final review: no Critical or Warning findings.

Closes #62060
Closes #62047