Enable multi-turn KV cache reuse

[stikves]

Summary

• Remove unconditional engine.reset() from CoreAIExecutor.respond() to let the KV cache persist across turns
• The engine's implicit prefix caching (TokenHistory.resolve) already handles all cases: same-prefix reuse, divergence rewinding, and full re-processing
• Saves re-prefill of shared conversation history on every turn

Local benchmarks (Qwen3 0.6B, M2 Max)

Multi-turn test with 2 prompts through LanguageModelSession:

	Prompt 1	Prompt 2
Before (full reset each turn)	0.5s	3.0s
After (cache reuse)	0.5s	2.4s

~20% improvement on prompt 2, scaling with conversation length (longer shared prefix = more prefill skipped).

Pipelined and sequential engines produce identical output with cache reuse enabled.

Closes #42

Test plan

• Full test suite passes
• Multi-turn FM API verification — both engine variants pass, outputs match
• LLM runner parity — determinism 3/3, quality 12/12, pipelined vs sequential match
• Verify engine.lastPrefixHitCount > 0 on turn 2+