feat(LlamaContext): expose kvUnified option

[andreinknv]

Summary

Adds a kvUnified?: boolean option on LlamaContextOptions, plumbing through to llama_context_params.kv_unified in llama.cpp.

When enabled, llama.cpp uses a single contiguous KV buffer indexed by sequence id rather than per-sequence buffers. On GPU backends (Metal, CUDA) this can significantly improve multi-sequence prefill/decode throughput by reducing per-sequence buffer juggling and enabling more efficient batched attention.

Why

llama-server already exposes this toggle as --kv-unified and most multi-slot deployments turn it on. Without a corresponding option in node-llama-cpp, callers running a single LlamaContext with sequences: N can't reach the same configuration, even when the underlying model + hardware would clearly benefit. The option was reachable via experimental*KvCache*Type for the dtype controls but not for the unified-buffer flag itself.

Plumbing

• src/evaluator/LlamaContext/types.ts — new kvUnified?: boolean option on LlamaContextOptions with full docstring (sibling to the existing swaFullCache field).
• src/evaluator/LlamaContext/LlamaContext.ts — _kvUnified member, constructor destructure, forwarded into the AddonContext options bag alongside swaFullCache, kvCacheKeyType, etc.
• llama/addon/AddonContext.cpp — when options.Has("kvUnified"), sets context_params.kv_unified to the unwrapped boolean.

Compatibility

• kvUnified is optional. When unset, the existing default ("whatever llama.cpp picks for this configuration") is preserved exactly — context_params.kv_unified is not touched.
• No public-API breakage. Existing callers see no behavior shift.

Test plan

• Local build + smoke test against Qwen2.5-Coder-3B-Instruct Q4_K_M (Metal, contextSize=8192, sequences=8) with kvUnified: true — context constructs cleanly; per-decode logs confirm kv_unified=1 reaches llama.cpp.
• Existing tests pass locally.
• CI on this PR.

🤖 Generated with Claude Code