docs: unified paged KV cache + disaggregated serving (B3c)

[inureyes]

Summary

B3c, the final sub-step of #126 (epic #116 capstone): benchmarks and documentation for the unified paged KV cache and the disaggregated serving roles. The correctness and parity tasks for #126 landed across the earlier sub-issues and the B-track PRs (#186, #187, #189, #190, #191, #192, #193); this PR adds the measured payoff and brings the docs in line, closing the epic.

What this does

• docs/turbo-kv-cache.md: a "Unified paged KV cache" section covering cross-request prefix reuse, copy-on-write block sharing, the --kv-cache-budget admission/eviction, and the GET /v1/cache/stats surface, plus a "Measured payoff" subsection with concrete numbers. States that paged adopt and donate are supported.
• docs/CONTINUOUS_BATCHING.md (new): the continuous-batching scheduler flags, how it uses paged decode, and the disaggregated prefill/decode/router roles (topology A) with the --node-role / --serving-bind / --prefill-peers / --decode-peers flags and a three-process localhost example.
• docs/README.md: index entry for the new doc.

Benchmark numbers (M1 Ultra, models/qwen3-0.6b-4bit)

• Memory: a shared 1024-token prefix is ~112 MiB of KV; across 8 concurrent requests the pool stores it once instead of eight times, saving ~784 MiB.
• Prefill avoided: an adopting request skips the shared prefill and decodes byte-identically to a cold run (tests/paged_prefix_share_parity.rs).
• Decode throughput (batch 4): the live native block-table kernel runs at 276 tok/s (512-token prompt) and 84 tok/s (4096), versus 146 and 7.7 tok/s for the gather reference (1.9x and 10.9x). The gather reference degrades sharply with context, which is why the live path uses the native kernel.

Notes

• docs/model_tests*.md and docs/en/prompt_cache.md named in the issue do not exist in the current flat docs layout; the unified-cache content lives in turbo-kv-cache.md.
• The issue's "remove the paged adopt/donate disabled note" task: no such note exists in the repo (it was never written), so the positive "supported" statement is added instead.
• Docs-only change; no code touched.

Closes #126