docs(predict-rlm): refresh RLM agent skills

[magix022]

Rationale

The previous split-skill PR (#35) became outdated while the PredictRLM runtime kernel and extension APIs continued to evolve. The installable guidance needs a fresh, source-accurate separation between building a callable RLM and optimizing an existing one, plus enough API and runtime context for an agent to design maintainable integrations.

Summary

• Rewrite the /rlm skill around the current PredictRLM execution model, callable API, signature instructions, reusable skills, files, tools, and runtime-kernel extension points.
• Add substantive signature and custom-Skill examples based on production RLM patterns and the built-in document skills.
• Add a dedicated /rlm-gepa skill for evaluation and instruction optimization of an existing RLM.
• Update the root and RLM-GEPA documentation to route users to the appropriate skill.

Test Plan

• Validate balanced Markdown fences and heading structure in all changed Markdown files.
• Parse every Python code block in both skill files and the updated READMEs.
• Construct the documented signature and custom Skill examples.
• Construct a custom adapter, RuntimeContribution, CallableTool, and EventSink through the documented kernel API.