Add Noodle Flights agent example with Phoenix tracing

[jabirabbas]

Adds a reference agent for the noodle_flights task suite, plus setup docs.

The agent drives the simulator with Playwright and uses Claude for the
reasoning steps (reading the form, identifying the cheapest option). It
currently passes all 8 noodle_flights categories.

Also wires up Arize Phoenix for observability: every browser step and LLM
call is traced, with a screenshot attached to each step span.

Files

• examples/noodle_flights_agent/agent.py — the agent
• examples/noodle_flights_agent/README.md — setup and usage

Notes

• No changes to any existing files; purely additive.
• Requires an Anthropic API key; Phoenix is optional (the agent runs
  without it, just without traces).
• Extraction-heavy categories read the flights API response directly and
  replicate the grader's filter_flights logic, since flight numbers
  aren't rendered in the results UI.

Happy to move or restructure this if examples/ isn't the right home.