Add session-aware prefix caching and /v1/stats endpoint - #199

Loaded 31944 chars (~10648 tokens) from The Project Gutenberg eBook of Crime and Punishment .txt

=== Test: Prefix Reuse (same session, same context) ===
  Q1 (12.38s): ...
  Q2 (13.23s): ...
  Q3 (13.58s): ...
  Prefix hits gained: 2 (expected ≥2)
  PASS

=== Test: Session Isolation ===
  New misses: 2 (expected ≥2)
  PASS

=== Test: Growing Context (multi-turn conversation) ===
  Turn 1 (3.27s): ...
  Turn 2 (3.26s): ...
  Turn 3 (3.27s): ...
  Turn 4 (3.27s): ...
  Prefix hits gained: 3 (expected ≥3)
  PASS

=== Test: Rapid Fire (20 short requests, sequential) ===
  20 requests in 2.2s (0.11s avg), errors: 0
  PASS

=== Test: Stats Coherence ===
  {'avg_prefill_tok_per_sec': 0, 'prefix_hit_rate': 0.819672131147541, 'prefix_misses': 11, 'total_gen_tokens': 861, 'total_requests': 61, 'avg_decode_tok_per_sec': 7.25221759620278, 'prefix_hits': 50, 'total_prompt_tokens': 77572}
  ✓ total_requests > 0
  ✓ prefix_hits + prefix_misses > 0
  ✓ 0 <= prefix_hit_rate <= 1
  ✓ total_gen_tokens > 0
  ✓ hit_rate matches hits/(hits+misses)
  PASS

==================================================
RESULTS:
  ✓ Prefix Reuse
  ✓ Session Isolation
  ✓ Growing Context
  ✓ Rapid Fire
  ✓ Stats Coherence

ALL PASSED