Add optional retry policy

For cases where the underlying LLM fails to generate valid structured data, or the data generated doesn't match certain post-response validation requirements, it may be desirable to retry the generation, prompting the LLM with feedback about what it did wrong, including the schema validation errors.

This may not always be desirable (it can introduce extra cost or latency), so it should be optional.

The retry policy should ideally be implemented at the Kurt level (not the adapter level) so that all adapters can benefit from it without needing to introduce it as a feature for each adapter.

Start at the Kurt-level generation flow rather than an individual adapter, focusing on where structured output and post-response validation failures are handled. Define what the optional retry policy receives and how validation errors are fed back to the LLM; done means retries are opt-in and apply consistently across adapters.

Written by the indexing model from the issue text.