Fix tool call ordering in transcript and add timestamps (#54)

[richardkiene]

Problem

The judge LLM was incorrectly detecting data fabrication because the conversation transcript showed the agent's text response BEFORE the tool call results.

What the judge saw (incorrect ordering):

[ASSISTANT]: The revenue is $1.5M based on the data...
  -> Tool call: get_revenue()
     Result: {"revenue": 1500000}

The judge thought: "The agent stated $1.5M before the tool returned that value - fabrication!"

Solution

1. Add timestamps to ToolCall model (called_at, responded_at)
2. Update ADK agent to capture precise timing of tool calls
3. Reorder transcript to show tool calls before response text
4. Add timestamps to all transcript entries

New format (correct ordering with timestamps):

[ASSISTANT] @ 10:30:45.123:
  -> Tool called @ 10:30:45.100: get_revenue({})
     Result @ 10:30:45.200: {"revenue": 1500000}
  Response: The revenue is $1.5M based on the data...

Now it's unambiguous that:

• Tool was called at 10:30:45.100
• Tool responded at 10:30:45.200
• Agent synthesized response using those results

Backwards Compatibility

• New called_at and responded_at fields are optional (float | None = None)
• Existing stored results load successfully (Pydantic uses None defaults)
• Timestamp formatter handles None by displaying "N/A"

Test plan

• All 255 unit tests pass
• Ruff linting passes
• Mypy type checking passes
• Backwards compatibility verified - won't break report generation

Closes #54