feat(mcp): Add per-tool-call metrics to `logs` response

[jaroslawgajewski]

Problem

The logs tool provides aggregate metrics per episode: total_tokens,
total_estimated_cost, mcp_failure_count. There is no breakdown by individual
tool call.

This means consumers cannot answer:

• "Which MCP tool consumes the most tokens?"
• "Which tool call failed and why?"
• "What is the latency distribution per tool?"

Current behavior

Only aggregates: total_tokens: 12840, mcp_failure_count: 1

Expected behavior

Include a tool_calls array per episode:

{
  "tool_calls": [
    {
      "tool": "get_file_contents",
      "server": "github",
      "tokens": 2400,
      "duration_ms": 350,
      "status": "success"
    },
    {
      "tool": "search_code",
      "server": "github",
      "tokens": 5200,
      "duration_ms": 1200,
      "status": "success"
    },
    {
      "tool": "create_pull_request",
      "server": "github",
      "tokens": 800,
      "duration_ms": 600,
      "status": "error",
      "error": "403 Resource not accessible by integration"
    }
  ]
}