fix: track last error across retries in _call_with_retry

[dennis-lynch-nv]

Problem

When all retries are exhausted in _call_with_retry(), the final RuntimeError only says:

LLM call failed after N retries for model X

...with no indication of what went wrong (rate limit? timeout? connection refused?).

Additionally, URLError exceptions are caught but not captured as a variable, so their details are lost.

Fix

• Add last_err tracking across all three exception handlers (HTTPError, URLError, TimeoutError/OSError)
• Capture URLError as e so the error details are available
• Include last_err in the final RuntimeError message
• Truncate HTTP body to 200 chars to keep error messages manageable

After

LLM call failed after 8 retries for model gpt-4o. Last error: HTTP 429: {"error":{"message":"Rate limit exceeded"...}}

Testing

Import-tested; no new dependencies. Change is limited to the retry loop in researchclaw/llm/client.py.