Improve HTML report rendering, tool call display, and fix empty schemas (#48, #49, #50)

[richardkiene]

Summary

• Render Judge LLM reasoning and Synthetic User messages as Markdown using the existing marked.js library
• Categorize tool calls into Required, Optional, and Unexpected sections with clear visual indicators
• Fix MCP server run_scenario returning empty tool schemas

Changes

Markdown Rendering (#48)

• Changed Judge LLM reasoning from plain text <p> tags to <div class="markdown-content"> for proper markdown rendering
• Changed Synthetic User (role="user") turns to use markdown rendering, same as agent under test output

Tool Call Categorization (#49)

• Rewrote _build_tool_calls_html to categorize tools by:
  • Required - tools that must be called (shows pass/fail status and missing tools)
  • Optional - tools from the scenario's optional_tools list
  • Unexpected - any other tools (highlighted with yellow background)
• Added optional_tools to the judgment result's tool_usage_dict so the report has access to it
• Added CSS styling for tool categories with visual differentiation

MCP Server Tool Schema Fix (#50)

• Added _extract_tool_schemas helper function that:
  1. First tries to extract schemas from mcp_server config (most reliable)
  2. Falls back to agent.get_available_tools() for ADK agents without mcp_server config
• This fixes reports generated via MCP server showing empty {} schemas

Test plan

• All 255 unit tests pass
• Ruff linting passes
• Mypy type checking passes
• Manual verification with ADK agent - tool schemas now display correctly

Closes #48
Closes #49
Closes #50