Adds a new `skills.promptMode` config option to control how skills
are injected into the system prompt:
- `full` (default): Current behavior - inject all skill metadata
- `compact`: Inject only skill names and truncated descriptions (~50% token reduction)
- `lazy`: Minimal prompt - skills available via list_skills tool (near zero overhead)
With 40+ skills installed, full mode uses ~3,700 tokens per request.
Compact mode reduces this by ~50%, lazy mode reduces to near zero.
Closes#3395
