perf(cli): use sliding window for download ETA calculation

[joohhnnn]

The ETA calculation uses cumulative average speed (total_downloaded / total_elapsed), which breaks after a resumable download retry: the downloaded counter jumps instantly to the resume offset while elapsed resets to zero, producing an infinite speed estimate and "ETA: 0s" when minutes of download remain.

Real data from a 239 GB snapshot download with 7 retries (71 min total):

After Retry	Progress	ETA Displayed	Actual Remaining
#2 (18.91%)	45.24 GB	0s	~57 min
#3 (39.49%)	94.48 GB	0s	~42 min
#4 (47.06%)	112.60 GB	0s	~37 min
#5 (64.27%)	153.76 GB	0s	~27 min
#6 (82.74%)	197.95 GB	0s	~16 min
#7 (92.36%)	220.98 GB	0s	~7 min

Root cause: after retry, DownloadProgress is recreated with downloaded = resume_offset but started_at = now, so speed = resume_offset / 0.001s = ∞.

Fix: replace cumulative average with window-relative speed that only counts bytes downloaded within the current 30-second window. On retry, window_downloaded is initialized to the resume offset so previously downloaded bytes are excluded from speed calculation entirely. This also simplifies the code into a single unified speed path and removes the now-unused started_at field.