Speed display was wildly inaccurate during parallel multipart uploads because:
1. Parallel chunk progress events caused non-monotonic byte counts
2. The 5-second smoothing window was too short for bursty uploads
3. Speed dropped to near-zero between chunk boundaries
Fixes:
- Enforce monotonic byte tracking via _peak guard
- Extend sliding window from 5s to 15s
- Cache last good speed value for display continuity
- Require >= 1s of data before calculating (was 0.5s)
