feat: dynamic skill loading for better context window

[nice-bills]

all available skills are injected into the system prompt on startup which consumes significant context tokens.

I propose moving to a "Just-in-Time" loading strategy:
Provide the agent with a high-level summary of available skills and add a tool like load skill details that would get the needed markdown
instructions only when the agent decides to use that skill.

I think this would improve performance on complex tasks and allow the skill library to scale a lot.