Attribute sidecar-only copies correctly and fix the remaining-time estimate

[Brandon-Haney]

Three fixes to what the completion banner and Recent Activity report.

Name the film when only its artwork was copied

Sidecars fold into their parent video's row, but the parent was indexed from the current run only. A video already on cache needs no copy, so a run that touched only its artwork produced unattributed rows: four ...-poster.jpg lines with nothing naming the film. The sibling map now covers every video with siblings rather than only the ones a run touched, so orphaned sidecars gather under a header carrying the video's filename in both surfaces. The header shows no size, because the video did not move and a size there would say it did. A lone sidecar stays flat, since its own filename already names the film.

Say when only a title's extras moved

That new header names the video, but it was still rendering the sidecars' own action badge, so a green CACHED sat beside a .mkv filename and read as though the video had been copied. FileActivity now carries sidecars_only through load, save and serialize, defaulting False so existing entries are unaffected. Both surfaces render a neutral tag instead: EXTRAS in the banner, "Extras only" in Recent Activity, each explaining in its tooltip that the video was already in place.

Price files as well as bytes in the remaining-time estimate

Copy time is bytes/rate plus a fixed per-file cost: open, size verify, chown, tracker write. The estimate modelled only the first, so a queue ending in artwork and NFOs read as nearly free. On a 199-file run it reported 9m 38s left when 11m 27s remained, because the 43 completed files averaged ~4 GB and the 156 still queued averaged ~0.75 GB.

_estimate_remaining_seconds() now splits the two. The best sustained rate seen is bulk throughput with per-file costs largely excluded, and whatever elapsed time that cannot explain is attributed to the files completed so far. Both are measured from the run in progress, so a slow array or a competing mover is reflected without a tuning constant. Sampled over 2s windows, since shorter ones are scheduling noise that would inflate the peak and hide the per-file term.

The banner also led with the file count while its bar and ETA were byte-based. At the moment above that read "43/199 files" beside a correct "9m left", which is a fifth against three fifths and makes the estimate look impossible. Bytes lead when byte progress is available, and the file count moves alongside elapsed.

How to test

1. Cache a title whose video is already on the pool but whose artwork is not. The banner and Recent Activity show one row naming the film, tagged EXTRAS / "Extras only", not a green CACHED beside the .mkv.
2. Confirm a single orphaned sidecar with no siblings still renders as a flat row.
3. Start a run with a mix of large videos and many small sidecars. The remaining-time estimate should stay stable as the queue shifts to small files instead of collapsing toward zero.
4. The banner leads with bytes transferred while byte progress is available, with the file count shown alongside elapsed.

Covered by tests/test_orphan_sidecar_grouping.py and tests/test_operation_eta.py.