Add loop_aware per-agent setting to skip context re-injection on loops > 0

Problem: SystemPromptSubscriber re-injects the full context block on every agent loop iteration. For agents with always_include items, this wastes ~22K tokens per loop — the LLM already has the context from loop 0 in its conversation window.

Solution: Add loop_aware boolean to ai_context.agents per-agent config. When enabled, SystemPromptSubscriber checks the current loop count (captured from AgentStartedExecutionEvent) and skips injection on loop > 0.

• Default FALSE: no behavior change for existing sites
• Missing key treated as FALSE: no update hook needed
• Per-agent granularity: single-loop agents always get context
• Admin UI checkbox added to the agent context form
• Config schema updated

Measured: 52% total token reduction on a heading edit operation (101K → 48K tokens, N=1, 15-component demo page with 8 ai_context items).

Closes #3582288