kv-cache: resolve TURBO_LAYER_ADAPTIVE mode per cache construction

[giveen]

Summary

Split out of #326 per review comment: #326 (comment)

adaptive_mode in llama_kv_cache::llama_kv_cache was a function-local static, so it was resolved once from whichever cache was constructed first in the process and then reused verbatim for every later construction, regardless of that cache's own type_v. This is a real bug for iswa/MTP setups that build more than one llama_kv_cache with different type_v in the same process, and for anything that sweeps --cache-type-v (e.g. llama-bench).

• Hoisted the mode resolution out of the per-layer loop into a per-construction local (kv_adaptive_mode), computed once per llama_kv_cache::llama_kv_cache call before the layer loop starts.
• Every layer within a given cache now sees the same mode (as before), and every new cache construction re-resolves the mode against its own type_v (the fix).
• No functional change to the mode-dispatch logic itself — this only fixes when the mode is resolved.

Test plan

• Builds clean (llama target)
• Single-cache-per-process behavior unchanged (mode still resolved once per construction)

🤖 Generated with Claude Code

[TheTom]

Merged. Per-construction resolution is the right shape; the second cache in an iswa or MTP process now gets its own mode instead of inheriting whichever was built first. The only visible change is that the auto-enable log line can print once per cache, which is fine.