[FEATURE] Re-architect skills

[edenreich]

Summary

Currently skills are 1-1 tools. But according to a more common established standard developed by Google, Anthropics and other providers skills are just markdowns that are being injected into the system prompt via metadata (frontmatter).
And the agent have a way to read them using a standard Read tool.

Tools should be called tools, and skills should be simply markdowns that describes different workflows so the agent is more aware of what it can do with the tools.

Skills should also be installable by the agent on startup - pull from the registry https://registry.inference-gateway.com/skills/

Acceptance Criteria

• Tools are renamed to tools - previously the agent schema (agent.yaml) was generating the tools, but those are not skills
• When defining skills in the schema it should be downloading the skills from the registry or if the user define bare to true the skill should be just a scaffold that the user can then write the skills themselves
• Each downloaded skill or created skill have frontmatter with metadata that is injected into the system prompt when skills are enabled.
• The skills defined in the schema are also the skills shown in the agent card - when the orchestrator fetches the card it should see them and not the tools.
• It's documented
• It's tested