fix(llm): enforce max_retries cap on the unbounded retry loop

[lfnothias]

Change

LLMProvider.__call__ retries in a while True loop. self.max_retries is set in __init__ but never read, so a persistently failing provider (timeout, overload, rate limit) retries indefinitely, re-sending the full prompt each time — unbounded latency, and unbounded spend where partial output is billed. Both retryable paths now raise once the attempts reach self.max_retries. The context-window path keeps its own independent cap.

Not changed

max_retries stays at 3. The success path and non-retryable errors are unaffected; a non-retryable error is still raised on the first attempt. The only change in behaviour is that a call which previously could not terminate now does.

Tests

tests/llm_provider_test.py: a persistent retryable error and repeated timeouts each raise after max_retries + 1 attempts; a non-retryable error raises on the first attempt. The network call and the sleep are mocked. Full suite: 198 passed, 4 skipped.