Refresh retrieval Q cache after single memory updates

[Frankie-Xu]

MemoryService.update_value persists a new Q value but leaves an already populated retrieval Q cache unchanged. A retrieve → single update → retrieve sequence can therefore keep ranking memories with the old Q value.

After a successful non-None updater result, refresh the cache using the same FIFO capacity policy as update_values. Disabled/null actions, None results, and persistence failures leave the cache unchanged. The Q-learning formula and retrieval policy are unchanged.

Validation:

• Five unittest regressions: two fail on main c1b322c and all pass with this change.
• The same five tests pass with pytest and from a clean git archive export.
• Tests import the real service, retrieval, normalization, batch update, and Q updater code, using in-memory persistence and fixed embeddings; only external MemOS import types are stubbed.
• compileall and git diff --check pass.

No live MemOS integration, LLM benchmark, or Python version matrix was run. This change is independent of the finite-value validation in PR #14.