Issue #3582288: Add loop_aware option to skip context re-injection on agent loops

Adds a loop_aware boolean to per-agent ai_context config. When enabled, SystemPromptSubscriber skips context injection on loop iterations > 0 because the LLM already has the context from the first loop in its conversation window. Reduces token consumption for multi-loop agents with large always_include context sets.

Includes 7 kernel tests with @dataProvider covering all loop_aware/loop combinations and the missing-key default case.

Closes #3582288