fix(turbokv): warn when --turbo-kv can't apply (e.g. with --ctx-size)

[solderzzc]

Problem

On mlx-swift-lm 460ff81, --ctx-size N gives the attention layers a RotatingKVCache (e.g. Qwen35 makeCache(capacity:)), and Server.swift only enables TurboQuant on KVCacheSimple layers. So --turbo-kv --ctx-size N logs turbo_kv=enabled and then silently compresses nothing. The M6 TurboKV numbers in #173 were really vanilla because of this (corrected in #180), and it hid #175 on the M6.

Fix: make it loud, don't change behaviour

• Startup: when --turbo-kv and --ctx-size are both set, print:
  ⚠️ --turbo-kv has no effect with --ctx-size N: a bounded context gives the attention layers a RotatingKVCache, and TurboKV only compresses KVCacheSimple. Drop --ctx-size to use --turbo-kv.
• Per request: the enable loop counts the layers TurboQuant was switched on for, and warns once if there were none. That also covers other models or cache setups where TurboKV can't attach.

Cache selection and TurboKV itself are unchanged. Making TurboKV work on RotatingKVCache is a separate decision, and while #175 stands (TurboKV loses long-range recall), a warning is the safer change.

Verified (Mac mini M6, Qwen3.8-27B-4bit, 2 requests per setup)

flags	warnings
--turbo-kv --ctx-size 16384	startup + per-request (printed once)
--turbo-kv	none
--ctx-size 16384	none

🤖 Generated with Claude Code

[solderzzc]

Reviewed (code-review skill, M5 agent): no issues found. I checked that --ctx-size goes through makeAttentionKVCache / makeHybridAttentionKVCache, which return RotatingKVCache when a capacity is set, so the startup warning holds for the models in the README. The per-request check also covers other cache types (e.g. CacheList). Merging.