feat(provider): interleaved reasoning replay

[xiami762]

Summary

This PR improves multi-turn reliability for reasoning / interleaved-thinking models and hardens the session runner against runaway tool loops.

Provider: interleaved reasoning replay

• Adds reasoning_replay.py and interleaved.py to normalize how assistant reasoning is echoed back to providers on follow-up turns.
• Extends Anthropic and OpenAI SDK paths to preserve reasoning_content / reasoning_details across tool-call turns, with provider-specific echo policies (when_present, tool_calls, all_assistant) and cross-provider promotion/placeholder handling.
• Updates catalog.json with interleaved-capability metadata for known reasoning model families.
• Integrates replay preparation into the session runner so multi-step tool flows keep valid reasoning context.

Session: tool-loop guards and step budget

• Introduces cross-step loop detection in SessionRunner:
  • Halt after 3 consecutive tool-only turns with the same tool + identical arguments
  • Halt after 8 consecutive tool-only turns using the same tool (even with different args)
  • Warn one step before each threshold
• Sets DEFAULT_MAX_TOOL_STEPS = 1000 when an agent does not declare an explicit steps limit (replaces unbounded inf).
• Message replay now resolves the model from the session’s current pin/default/agent override, not the historical model stored on the original user message.

Agent: safer plugin loading

• Hardens agent.yaml loading: validates mapping shape and skips malformed entries without breaking agent discovery.