Add `RetryEvaluator`

[ptomecek]

Add a generic retry evaluator for wrapping CallableModel invocations that may fail transiently (flaky APIs, databases, object stores, etc.).

This pattern is currently re-implemented ad-hoc by users. A first-class evaluator would standardize the behavior and compose cleanly with the rest of the evaluator stack.

Behavior

• Configurable maximum number of attempts.
• Configurable backoff policy (at minimum: constant and exponential, with optional jitter).
• Configurable allow-list and deny-list of exception types to retry on.
• Optional hook invoked between attempts (for logging, metrics, etc.).

Open questions

• Idempotency. Retries assume the wrapped model is safe to re-invoke with the same context. Should CallableModel gain an explicit idempotency flag that this evaluator respects, or should it be the user's responsibility to only stack retry on idempotent models?
• Interaction with caching. When stacked with MemoryCacheEvaluator, transient failures should not be cached. Need to confirm or adjust the cache evaluator's behavior on exceptions, and document the canonical stacking order.