Upload UX polish: per-chip status, done state, size-aware ETA; fix gear position

[aperson30]

Summary

• Removed the page-wide "Uploading ct.nii.gz..." status line. Each file chip in the dropzone now shows its own status (spinning ring while uploading, green check when done, red tint on failure) instead - the whole dropzone box turns green once every selected file has finished uploading.
• Restyled file chips (white card + shadow) so they read as distinct objects sitting in the dropzone instead of blending into its near-identical gray background.
• The inference ETA now scales with file size (previously one flat number per model, off by a lot on anything not close to the one case it was calibrated on) and prefers a real server-measured median once the backend has history for that model/size - GET /api/inference-duration-estimate. The backend now logs (model, file_size_bytes, duration_seconds) for every completed job (measured from GPU-slot grant, not queue entry) to build that history; falls back to the size formula silently when there isn't enough data yet.
• Viewer: the floating gear shown when the toolbar is hidden was positioned at a fixed left offset that landed in the middle of the axial pane's own CT image on a normal-width screen - anchored to the top-right corner instead.

Also found while investigating "ePAI only shows a few organs"

Traced this end-to-end - the website's label mapping and viewer code (_EPAI_TO_VIEWER, segmentation_categories, OrganCheckbox) were already correct. The real bug was in the ePAI model's own export pipeline (~/ePAI/binary/nnunetv2/inference/export_prediction.py on the GPU server - a separate repo I don't have push access to, same situation as an earlier session's CUDA-cache fix): an explicit "extract only pancreas and tumor" step was zeroing out every other organ before the segmentation file was even saved, regardless of what the model detected. Verified via a real inference run: before the fix, only 2 of 25 possible labels were ever written to disk; after removing the masking step, 21 labels came through on the same test case, with pancreas/tumor stats unchanged within normal GPU nondeterminism. Fixed live on the server and the warm predictor restarted to pick it up - not part of this PR's diff since it's a different repo, documented here for visibility.

Test plan

• npm run build (tsc -b + vite build) - clean
• vitest run - 298/298 passing
• Verified the ePAI export fix directly against a real test case (before: labels {0, 17, 22} only; after: 21 distinct labels, matching the model's full 25-class training set)
• Manual: select multiple files, confirm each chip shows uploading -> done and the box turns green
• Manual: run ePAI on a real scan, confirm the ETA text updates plausibly and the viewer now shows more than pancreas/tumor