fix(perf): ignore metadata-only chunks in TTFT and ITL - #1645

Problem

OpenAI-compatible chat streams can emit choices events without generated output. For example,
vLLM sends an initial role-only delta with content: "", and services may send a separate empty
delta carrying finish_reason.

DefaultApiPlugin.process_request() currently records TTFT for the first event with a non-empty
choices array and records every later choices event as an ITL sample, even when the extracted
text and reasoning content are empty. This can report TTFT before the first visible output and add
metadata gaps to ITL.

For a deterministic stream with events at 0.10s (role), 0.45s (H), 0.65s (i), and 0.90s
(finish), the current code reports:

• TTFT: 0.10s
• ITL: [0.35s, 0.20s, 0.25s]

The output-based result should be:

• TTFT: 0.45s
• ITL: [0.20s]

Fix

Track the last non-empty output timestamp separately from the last response-event timestamp.

• Update TTFT and ITL only when text or reasoning content is non-empty.
• Keep the existing response timestamp for completion time and end-to-end query latency.
• Continue retaining metadata events and parsing usage from them.

Verification

The new asynchronous regression reproduces role-only, content, finish-only, usage, and [DONE]
events with deterministic timestamps. Before the fix, it fails with 0.10 != 0.45 at the TTFT
assertion. After the fix:

• 31 related offline tests pass on current main.
• All configured pre-commit hooks pass for the two changed files.
• compileall and git diff --check pass.

The regression also verifies that generated text, response retention, usage parsing, and
end-to-end query latency remain unchanged.