Add producer/critic reflection loops to CustomToolAgent

Builds on #22615's integrated UserToolSource validation. The producer
agent runs with output_retries=0 so the reflection loop owns the retry
and can pass a structured error list back to the prompt rather than
pydantic-ai's generic 'try again' message.

Two opt-in loops on top of pydantic validation:

- Validator-driven retry (default on): if the producer's output fails
  validation, _produce_tool returns the formatted error list, and
  process() re-calls the producer once with those errors prefixed to
  the prompt. Cap of one retry; if it still fails, the agent returns
  a low-confidence validation_failed response.
- Quality critic + refine (default off): an LLM critic agent reviews
  the validated tool for clarity (description, labels, help text) and
  idiomaticity (defaults, exposed options, container choice) -- the
  fuzzy dimensions pydantic can't see. If the critic flags significant
  issues (should_refine=true), the producer is re-rolled once with
  the critique. Cap of one refine; if refinement breaks validation,
  the original tool is kept rather than ship something worse.

Both gates live under inference_services.custom_tool:
validator_retry_enabled (default true) and quality_critic_enabled
(default false). Defaulting the critic off keeps the cost neutral for
deployments that don't opt in -- the critic doubles tool-creation
latency / spend when enabled.

The 170-line process() method got factored into named helpers
(capability check, structured-output extraction, validation-failed
response, success response, model-error handlers) to keep the
reflection control flow readable.

Author: dannon