fix(agent): align AgentMemory decay with EbbinghausAlgorithm

[knqiufan]

Summary

• Replace the hard-coded linear decay in MultiAgentMemoryManager.update_memory_decay() and MultiUserMemoryManager.update_memory_decay() with EbbinghausAlgorithm.calculate_current_retention(), keeping Agent cache retention_score in sync with metadata.intelligence.
• Introduce shared helpers in src/powermem/agent/utils/retention.py for retention calculation, cache/metadata synchronization, cleanup classification, and persistence.
• Update cleanup_forgotten_memories() in both managers to use Ebbinghaus should_forget() / should_archive() instead of fixed linear thresholds on stale cache scores.
• Persist decay/archive metadata through storage.update_memory() so scheduled decay passes do not re-trigger the intelligence plugin.
• Preserve legacy string memory IDs when loading memories from storage into Agent caches, while retaining normalized int/string cache variants for modern Snowflake IDs.

Test plan

• python -m pytest tests/unit/test_issue_925_agent_memory_share.py::test_multi_agent_loads_legacy_agent_scope_as_agent_group -q
• python -m pytest tests/unit/agent/test_memory_id_normalize.py -q
• python -m pytest tests/unit/agent/test_agent_retention_decay.py -q

Fixes #1149

[wayyoungboy]

LGTM