feat: LoopAgent exit_condition, ReflectRetry plugin, RunConfig, weather example

[zeroasterisk]

Changes

1. LoopAgent enhancement (#89) — P1

• Added exit_condition function field that receives context after each iteration
• Context temp_state merges between iterations for state propagation
• Graceful stop when no exit_condition and max_iterations reached
• 15 thorough tests: basic loop, early exit, max iterations, nested loops, escalation interaction

2. ReflectRetry Plugin (#78) — P1

• New ADK.Plugin.ReflectRetry that intercepts error events in after_run/3
• Feeds errors back to the agent with reflection context ("try again, here's what went wrong")
• Configurable max_retries (default: 3)
• Uses existing Plugin + Callback system
• 8 tests covering init, error detection, retry success, max retries exhaustion

3. RunConfig (#90) — P2

• New ADK.RunConfig struct: streaming_mode (:none/:sse/:live), max_llm_calls, output_format, speech_config
• Validation on construction (invalid streaming modes, non-positive max_llm_calls)
• Integrated into Runner (accepts run_config: option) and Context
• 12 tests + 2 doctests

4. Weather Agent Example (#93 partial)

• Complete runnable Mix project in examples/weather_agent/
• Demonstrates agent definition, tool creation, runner usage, interactive multi-turn conversation
• 8 tests, README included

Test Results

• All new tests pass (37 tests, 4 doctests, 0 failures)
• Weather agent example: 8 tests, 0 failures
• Pre-existing failures unchanged (SessionSupervisor/Registry not started in test env)