feat: expose paged cache serving settings

[0z5a]

Summary

• expose paged KV cache token capacity and page size through the serving CLI and YAML config
• propagate the settings through the HTTP server, engine, and scheduler while preserving contiguous defaults
• validate invalid capacity/page combinations and report effective cache mode through /stats
• document CLI, YAML, Docker, paged allocation, and prefix-caching behavior

L20 capacity benchmark

Exact PR commit 323f907, NVIDIA L20 (SM89), BF16, PyTorch 2.11.0+cu128, CUDA 12.8. The allocation uses the real AstrAI 1.2B KV geometry: 24 layers, 4 KV heads, head dimension 64, maximum batch 16, and maximum sequence length 32,768.

serving configuration	reserved token slots	page size	prefix cache	CUDA allocation	cold init
default contiguous	524,288	1	off	12.002 GiB	13.04 ms
kv_cache_tokens=131072, kv_cache_page_size=64	131,072	64	on	3.002 GiB	5.05 ms

For a workload that needs four rather than sixteen simultaneous full 32k contexts, the exposed capacity setting saves exactly 9.0 GiB of KV-cache allocation (about 75%). This is a capacity/configurability result, not a claim that paged attention itself is 75% faster; throughput still depends on prompt sharing, occupancy, and backend.

GET /stats makes the effective mode, capacity, page size, and prefix-cache state visible so deployments can verify that the requested setting took effect.

Validation

• ruff format --check .
• ruff check . --select I
• python -u -m pytest tests/ -q (664 passed)