[codex] fix prefix cache for user-first prompts

[easel]

What changed

• Allows prefix-cache boundary detection to start from the first chat role marker when a prompt has no system message.
• Emits usage.timings.prompt_n_cached when the server restores a cached prefix.
• Adds model-free unit coverage for Qwen-style system-first and user-first prompts plus timing telemetry.

Why

Agentic multi-turn prompts commonly start with a user message. The prefix cache previously required the system-role marker before it would identify boundaries, so user-first sessions never activated the cache even when subsequent turns shared a long prefix.

Validation

• cmake --build build --target test_server_unit dflash_server -j 8
• ./build/test_server_unit