feat(inference): add tool schema translation layer

[Svagtlys]

feat(inference): add tool schema translation layer

Description

Build a translation layer that converts MCP tool schemas (inputSchema in JSON Schema format) into the provider-native tools array format expected by the inference engine (e.g., OpenAI's function schema or Anthropic's tools parameter).

Why

MCP servers expose tools in MCP's JSON Schema format. The inference provider (OpenAI-compatible) expects a different schema structure. The harness must translate between the two so the LLM can understand the available tools when making completion requests.

Scope

• Add tools field to CompletionRequest type to carry tool definitions
• Build translator: MCP inputSchema → OpenAI-compatible function.parameters format
• Pass translated tool definitions alongside messages in completion requests
• Support multiple tool definitions in a single request

Depends on: #8, #15