Add TwoSpeedEngine: fast/slow inference routing with embedding cache

Implements the two-speed inference pattern from TWO_SPEED_ARCHITECTURE.md
for Phase B (single-node mode):

- TwoSpeedConfig: tunable thresholds for routing, caching, and decoding
- ComplexityEstimator: scores queries via length + embedding-space novelty
- EmbeddingCache: LRU cache of latent plans (OrderedDict, cosine-sim lookup)
- FastPath: decodes directly from the nearest cached latent plan
- SlowPath: rolls out VLJEPA dynamics for deep reasoning then decodes
- NetworkStub: Phase B shim that routes slow-path calls to local rollout();
  interface designed for drop-in replacement with a real network client in Phase C
- TwoSpeedEngine: orchestrates routing, fallback (fast→slow on low confidence),
  cache population, and decoding; exposes both async infer() and sync infer_sync()

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Author: EightRice