feat(perception): implement memory-first image storage strategy

[TexasOct]

Implement a memory-first storage strategy to reduce disk I/O and optimize disk usage.

Key Changes:

• ImageManager: Add memory-first mode with dynamic TTL calculation

  • New methods: persist_image(), persist_images_batch(), cleanup_batch_screenshots()
  • TTL-based cleanup for memory-only images (default: processing_interval * 2.5)
  • Track image metadata (timestamp, is_persisted) for lifecycle management

• ActionAgent: Trigger persistence only when actions are created

  • Persist screenshots to disk before saving actions
  • Immediate batch cleanup for unused screenshots after action generation

• Coordinator: Add periodic TTL cleanup in processing loop

• Config: Add memory-first configuration options

  • enable_memory_first, memory_ttl_multiplier, memory_ttl_min/max

Benefits:

• Reduce disk writes by ~95% (from 3,600/hour to ~150/hour)
• Immediate cleanup of unused screenshots (vs. waiting for TTL)
• Dynamic TTL based on processing_interval for adaptive behavior

Three-tier cleanup mechanism:

1. Batch cleanup: Remove unused images immediately after action creation
2. TTL cleanup: Evict expired memory-only images every processing cycle
3. LRU eviction: Fallback when cache limit (500 images) is reached