feat: support inline skill references with $ autocomplete

[somewan820]

Problem

Pi currently exposes loaded skills as /skill:name slash commands. Slash command completion is only available at the beginning of a message, so a user cannot reference a skill naturally inside an already-written prompt.

For example, this does not offer skill completion:

Please review this endpoint using $farbay-java-develop

Proposal

Add inline skill references using $skill-name:

• Typing $ at a token boundary opens completion for the skills currently loaded in the session.
• Tab accepts a skill reference without submitting the message.
• On submit, the selected reference is expanded to the native skill content before the model turn.
• $ references work anywhere in the user message.
• Existing /skill:name behavior remains unchanged.
• Existing @path/to/file completion and file-reference behavior remains unchanged.

Example:

Please review this endpoint using $farbay-java-develop, focusing on authentication.

Implementation constraints

• Use Pi as the source of truth for loaded skills, resource loading, trust, and reload lifecycle.
• Do not duplicate the skill loader in OpenPI.
• Keep the raw user-visible message distinct from the expanded model-visible prompt where the existing lifecycle supports that distinction.
• Define behavior for unknown skills, escaped dollar signs, multiple references, and references adjacent to punctuation.

Acceptance criteria

• Inline $ completion lists only currently loaded skills.
• Selecting a skill does not submit the message.
• A submitted inline reference actually loads the corresponding SKILL.md content.
• Existing slash skill commands and @ file completion do not regress.
• Tests cover completion, expansion, unknown references, and mixed inline text.

[bkyssn]

I'd like to work on this issue.