feat: add MemoryCount, score normalization, and dedup-aware MemoryStore

[daniel-harperdb]

Summary

Adds the three features that were lost during stacked PR squash-merges (#7, #8, #9):

• MemoryCount — POST /MemoryCount returns count of memories with optional filters (source, classification, channelId, authorId, agentId)
• Score normalization — MemorySearch and SynapseSearch now return a similarity score (0–1) alongside $distance for LangChain compatibility
• MemoryStore with dedup — POST /MemoryStore with two-phase deduplication: SHA-256 content hash for exact matches, vector similarity for semantic near-duplicates. Adds contentHash and supersedes fields to schema.

Test plan

• test/memory-count.test.js — 6 tests for count with various filters
• test/score-normalization.test.js — 8 tests for MemorySearch and SynapseSearch normalization
• test/dedup-store.test.js — tests for exact match, fuzzy match, and new record paths