feat(bench): implement per-scenario memory isolation (Qdrant + SQLite)

[bug-ops]

Description

Before each benchmark scenario begins, reset the bench-session Qdrant collection and SQLite database so earlier scenarios cannot contaminate later ones.

Part of epic #2827. See spec: .local/specs/zeph-bench/spec.md NFR-001, FR-002.

Scope

• BenchIsolation::reset async function that:
  • Deletes and recreates the Qdrant collection named bench_<dataset>_<run_id>
  • Clears or recreates the SQLite conversation-history tables for the bench session DB path (bench-<run_id>.db)
• Bench session config overrides: memory.qdrant.collection and DB path are set to bench-namespaced values before agent construction
• Verify production collection names (zeph_memory, zeph_skills, etc.) are never touched
• Isolation overhead must be < 2 seconds (NFR-007)

Acceptance Criteria

• After scenario N completes, no memories from N are visible to scenario N+1
• Production Qdrant collections and SQLite DB are untouched after a bench run
• Isolation reset completes in < 2 seconds on a local Qdrant instance
• Integration test (ignored, requires Qdrant) verifies isolation