fix(memory): use effective_embed_provider for all embedding ops in semantic submodules

[bug-ops]

Summary

• Replaces self.provider with self.effective_embed_provider() at all 18 embedding call sites across recall.rs, cross_session.rs, summarization.rs, and corrections.rs in zeph-memory
• Adds regression test effective_embed_provider_routes_to_dedicated_embed_provider that verifies the dedicated embed provider is used instead of the main LLM provider

Root Cause

self.provider (main LLM provider, e.g. OpenAI/text-embedding-3-small, 1536 dim) was used at recall time instead of self.effective_embed_provider() (e.g. Ollama/nomic-embed-text-v2-moe, 768 dim). ensure_collection() detected the dimension mismatch and silently deleted and recreated all Qdrant memory collections on every session startup, losing all stored embeddings.

Test Plan

• cargo build -p zeph-memory — passes, 0 errors
• cargo clippy -p zeph-memory -- -D warnings — 0 warnings
• cargo nextest run -p zeph-memory --lib — 1115 passed, 7 skipped, 0 failed
• New regression test effective_embed_provider_routes_to_dedicated_embed_provider passes and would fail if any fixed call site reverted to self.provider
• Verify with mixed-provider config: zeph_conversations collection no longer recreated on startup

Closes #3029