fix: measure token speed from streaming phase only (exclude TTFB)

[kianwoon]

Summary

• Token speed (TOK/S) previously used total wall-clock latency (including TTFB wait), producing misleadingly low averages
• Now captures _streamStartTime when the first streaming chunk arrives, and computes TPS using streaming-only duration
• Adds a 200ms minimum duration threshold — falls back to total latency for fast responses where timing precision is unreliable (prevents inflated numbers like 64K tok/s)

Changes

• src/types.ts — added _streamStartTime?: number to RequestContext
• src/server.ts — capture streaming start on first chunk (SSE + JSON paths), use streaming-only duration in recordMetrics with 200ms floor

Test plan

• npx tsc --noEmit — passes
• npm run build — succeeds
• npx vitest run — 323/323 tests passing
• Manual: verify TOK/S shows realistic values in GUI for long streaming responses (should be higher than before)
• Manual: verify short/fast responses don't show inflated numbers