fix(channels): catch all TimeoutException subclasses and safely parse Retry-After in retry_request

[tejajakarulloh]

Fixes #642

Problem

retry_request (src/agentos/channels/_util.py) has two bugs:

1. Incomplete timeout handling — catches (httpx.ConnectError, httpx.ReadTimeout) but misses ConnectTimeout, WriteTimeout, and PoolTimeout. These are all httpx.TimeoutException subclasses, not ConnectError subclasses (separate inheritance branches under TransportError). A DNS/TLS handshake timeout crashes through without retry.

2. Unsafe Retry-After parsing — float(resp.headers.get("Retry-After", ...)) raises ValueError if the header contains an HTTP-date string (RFC 7231 §7.1.3 permits both formats).

Fix

• Widen the exception clause from (httpx.ConnectError, httpx.ReadTimeout) to (httpx.ConnectError, httpx.TimeoutException).
• Wrap Retry-After parsing in try/except (ValueError, TypeError) with fallback to exponential backoff; clamp with max(0.0, ...).

Testing

New test file tests/test_channels/test_channel_retry_util.py covers:

• All four TimeoutException subclasses are retried (ConnectTimeout, ReadTimeout, WriteTimeout, PoolTimeout)
• ConnectError still retried (unchanged behavior)
• Exhaustion after max retries raises
• Numeric Retry-After header honored
• Non-numeric / HTTP-date / missing Retry-After falls back to backoff

All existing channel tests pass (306 passed).