server: add KV cache metrics

[lvsijian8]

Fixes #23632.

Problem

The Prometheus /metrics endpoint exposes request and throughput metrics, but it does not report KV cache pressure.

Changes

• add a read-only KV cache cell usage query on llama memory
• report KV cache used cells, total cells, and usage ratio from the server metrics task
• document the new Prometheus gauges
• cover the /metrics output in the server basic tests

Tests

• .venv/bin/cmake -B build
• .venv/bin/cmake --build build --target llama-server -j
• PATH="$PWD/.venv/bin:$PATH" LLAMA_CACHE="$PWD/tools/server/tests/tmp" PORT=18080 LLAMA_SERVER_BIN_PATH="$PWD/build/bin/llama-server" ./tools/server/tests/tests.sh unit/test_basic.py::test_server_metrics_kv_cache -q