fix(copilot): consolidate agent_message_chunk events into single log lines

[christso]

Problem

When running agentv eval with a Copilot target, the stream log files are hard to read because each agent_message_chunk ACP event is written as a separate log line. A single assistant response can produce dozens of fragmented lines:

```
[+00:02] [agent_message_chunk] Let me
[+00:02] [agent_message_chunk] analyze
[+00:02] [agent_message_chunk] this code
[+00:02] [agent_message_chunk] and fix the bug
```

Root Cause

CopilotStreamLogger.handleEvent() in copilot-utils.ts emits one log line per ACP event. The summarizeAcpEvent() function in copilot-cli.ts limits chunks to 200 chars each, but does nothing to consolidate them.

Best Practice (from code-insights research)

code-insights uses a flush-on-turn-boundary pattern:

• Accumulate assistant.message_delta / chunk events into a text buffer
• Flush the buffer as a single complete entry only at turn boundaries (tool_call, session.idle, session.shutdown, or close())
• Result: [+00:05] [assistant_message] Let me analyze this code and fix the bug...

Proposed Fix

Modify CopilotStreamLogger in copilot-utils.ts:

• Add a pendingText buffer for chunk accumulation
• In handleEvent(): buffer chunk events instead of writing immediately; flush on non-chunk events
• In close(): flush any remaining buffered text before closing

Purely a log formatting improvement — no behavioral change to eval execution.

Acceptance Criteria

• Single assistant response produces one [assistant_message] log line, not dozens of [agent_message_chunk] lines
• Tool calls, session start/end still appear as separate lines
• json format mode is unaffected (keep per-event for full fidelity)
• Both copilot-cli and copilot-sdk providers benefit (shared CopilotStreamLogger)
• Unit tests updated