fix: recover from transient reply failures

[FrundlesTian]

Description

Prevent intermittent Telegram requests from finishing without a delivered reply.

• retry transient Telegram NetworkError/TimedOut failures with bounded exponential backoff
• keep invalid-markup fallback separate so permanent BadRequest failures are not retried
• make typing indicators and progress-message deletion non-fatal
• retry progress, final-response, classic-handler error, and global error messages
• require the Claude SDK stream to end with ResultMessage
• retry only a completely empty SDK stream; do not replay a partially observed stream because tool side effects may already have happened
• configure the separate getUpdates client with the explicit proxy

This is stacked on #236 and currently includes #240 (4f62dc2). The connection-stability implementation is commit 0b8c8e2; once #240 lands, this branch can be rebased without changing the fix.

Related issue

Closes #241

Type of change

• Bug fix
• New feature
• Breaking change (documented in CHANGELOG under "Changed" or "Removed")
• Documentation or tooling only

How it was tested

• Tests added or updated
• pytest -q — 690 passed
• black --check src tests
• isort --check-only src tests
• flake8 src tests
• Tested by hand against a running bot

The regression tests cover transient recovery, retry exhaustion, permanent BadRequest, empty SDK EOF recovery, partial-stream non-replay, and proxy wiring for both PTB clients.

Checklist

• One concern per PR; unrelated changes are split out
• CHANGELOG.md has an entry under [Unreleased]
• No dependency changes
• No settings or command documentation changes required
• New settings default to current behaviour (no new settings)
• AI-assisted changes were reviewed line-by-line; no end-to-end bot claim is made

[github-actions]

No issues found that should block merging.