feat(core): dual-mode agent loop for native tool_use

[bug-ops]

Problem

Agent loop only supports text-based tool extraction. Need to branch based on provider capability.

Solution

In process_response():

• If provider.supports_tool_use(): use chat_with_tools(), handle ChatResponse::ToolUse
• Else: use existing text extraction path (bash blocks)
• Tool execution reuses same ToolExecutor pipeline
• ToolResult fed back as ToolResult MessagePart

Acceptance Criteria

• Agent detects provider tool_use support
• Native path: structured tool calls → execute → tool result → next iteration
• Text path: unchanged behavior
• Both paths respect PermissionPolicy
• Both paths respect max_tool_iterations and doom-loop detection

Part of #253 (M20)