Feature enhancements: thresholding, train/prediction, 3d volume

[davramov]

Summary of what is in this PR:

• 3D volume viewer + Zarr/TIFF support: integrated the vendored WebGPU renderer, TIFF-stack loading alongside Zarr, higher-resolution mask labels in 3D, viewing predicted regions directly in the volume.
• Threshold tooling: added a blur control, an auto-threshold "eyedropper" magnetic-lasso tool, and a fix for clipping on threshold-painted slices (plus general Annotate-tab speedups).
• Denoising: Alex's classical + learned denoising work added to the Annotate tab.
  dlsia deep training (new Train tab): the full training/inference pipeline ("Alex's deep training," end-to-end tested), cross-slice iPred support, a "Deep" mask layer independent of iPred's "Fast" layer, live per-slice preview during inference, a fix for iPred/dlsia writes clobbering each other in Tiled, and splitting "Push to Tiled" from "View in 3D" into independent actions.
• iPred (fast in-browser classifier) integration: full feature-bank/manifold-sampling/prediction pipeline, plus UI polish (collapsible panels, denoising-matched styling).
• Docker/CI: updated Docker images and start_all.sh hardening.
• Test coverage: backend 46%→92%, frontend RTL introduced and driven 19%→74%.
• Bug-fix chain: the mask cache-key collision, the missing "Write masks to Tiled" button, a numpy IndexError crash, a false-timeout on mask loading, and Peter's bucket-fill class-boundary fix — plus MkDocs documentation updates.

Next steps:

• Working on deployment hardening