feat(llama-cpp): make server-side prompt cache work by default

[localai-bot]

Summary

Fixes #9921. Repeated system prompts (agents, OpenAI/Anthropic-compatible CLIs, coding assistants) now hit the server-side prompt cache out of the box: no YAML changes required.

LocalAI's llama-cpp gRPC backend hardcodes `n_parallel=1` and was also hardcoding `kv_unified=false`, which silently force-disables `cache_idle_slots` at upstream's server init (server-context.cpp:1023-1034). The host prompt cache was being allocated (`cache_ram_mib=-1` = no limit) but never written across requests, so a 20-25k-token system prompt got re-prefilled on every call. Upstream sidesteps this by bumping `n_parallel` to 4 and flipping `kv_unified=true` when the user leaves slot count on auto (tools/server/server.cpp:100-105); LocalAI never hit that path.

This PR:

• Flips `kv_unified` default to `true` in `backend/cpp/llama-cpp/grpc-server.cpp` (keeping `n_parallel=1` to avoid a slot-count behavior change). Single-slot setups now get prompt-cache hits across requests.
• Bumps `n_ctx_checkpoints` default from 8 to 32 to match upstream.
• Initializes `cache_idle_slots=true` and `checkpoint_every_nt=8192` explicitly (both match upstream defaults).
• Exposes `cache_idle_slots` / `idle_slots_cache` and `checkpoint_every_nt` / `checkpoint_every_n_tokens` as new option keys so users can opt out or tune.
• Fixes docs: the `cache_ram` description was wrong (it's the host-side prompt cache, not the KV cache). Documents the kv_unified + cache_ram + cache_idle_slots interaction, adds rows for the two newly-exposed options, and adds an explainer worked example for the repeated-system-prompt workload from the issue.
• Marks the legacy `prompt_cache_path` / `prompt_cache_all` / `prompt_cache_ro` YAML fields as unused by the llama-cpp gRPC backend in `docs/content/advanced/model-configuration.md` (they target upstream's CLI completion tool and are not read by `grpc-server.cpp`) and point readers at the new prompt-cache explainer.

Test plan

• `make -C backend/cpp/llama-cpp clean && make backends/llama-cpp` builds clean (done locally)
• Smoke test: load any GGUF, send the same 5k+ token system prompt 3 times via `/v1/chat/completions`. Call 1 prefills; calls 2-3 should hit the warm prompt cache and finish in seconds. Check server log for `prompt cache is enabled` and `idle slots will be saved to prompt cache and cleared upon starting a new task`.
• Verify `options: ["kv_unified:false"]` restores the old behavior (no idle-slot saving, prefill repeated).
• Verify `options: ["cache_ram:0"]` disables the prompt cache entirely.
• Verify `options: ["cache_idle_slots:false"]` keeps `kv_unified=true` but disables idle-slot saving.

Reported and analyzed by @pos-ei-don on a DGX Spark (GB10/ARM64, CUDA 13): 5-8 min per turn for Claude-Code-style sessions without the cache, collapsing to seconds with it on.

🤖 Generated with Claude Code