feat(inference): pin each conversation to its own llama-server slot

[ganisback]

Summary

• With two or more parallel slots, each conversation stays on one llama-server slot, so agent sessions that share a system prompt stop evicting each other's prompt cache.
• The conversation key now covers the whole first turn, plus Anthropic system and Responses API input, so sessions that share a preamble stay distinct across chat, Claude, and Codex requests.
• A single slot is unchanged and does not send id_slot.

Test plan

• Set parallel slots to 2 or more, run two agent sessions on the same local model, and confirm each keeps its prompt cache across turns.
• With one slot, confirm requests do not send id_slot.
• Confirm a follow-up tool turn stays on the same slot, and two sessions that share a system preamble get different slots.
• Check the parallel-slots setting description in English and Chinese.