Add compact skills prompt mode for small-context models

[bug-ops]

Parent: #740 (P1)

Problem

Full SKILL.md inlining in system prompt wastes context budget on models with small context windows (local Ollama, Candle). A single skill can consume 1-2K tokens.

Solution

• Add compact prompt mode: inject only skill name + description + file path
• Agent reads full instructions on-demand via a read_skill tool call
• Auto-select mode based on model context window size (threshold configurable)
• Manual override via skills.prompt_mode = "full" | "compact" | "auto"

Affected crates

• zeph-skills (prompt generation)
• zeph-core (mode selection, optional read_skill tool)

Acceptance criteria

• Compact mode reduces skill prompt tokens by >70%
• Auto-selection based on model context size
• Full instructions retrievable on demand
• Config override works