Add SPIKE-618 LLM tool-calling architecture doc

Add completed spike document for SPIKE-618 describing the proposed LLM tool-calling architecture for Taskdeck chat. Covers decisions (custom implementation vs Semantic Kernel), provider abstraction and TaskdeckToolSchema, ILlmProvider extension (CompleteWithToolsAsync), ToolCallingChatOrchestrator multi-turn loop, read/write tool inventory and schemas, mock provider strategy, SignalR status streaming, persistence of tool-call metadata, cost/latency model, risks, and a phased implementation plan to guide development and testing.

Author: Chris0Jeky