Drop mean std

[schilling40]

Drop the mean/std step

The cochlea-net pipelines no longer compute a global mean and standard deviation before the prediction. The first step of the SGN, IHC and synapse pipelines now only creates mask.zarr.

Why the step can go

Every production U-Net of cochlea-net (synapse AnisotropicUNet, SGN/IHC UNet3d) starts with InstanceNorm3d(affine=False). This layer standardizes each input block by the mean and std of the block itself. A global (x - m) / s is one affine map for the whole block, so the layer cancels it. Only eps and float rounding remain. The measured relative difference of the predictions is 6e-6 to 2e-5. The later steps read only predictions.zarr and mask.zarr. Each run saves one pass over the volume.

Changes

• The input guards of apply_SGN, apply_IHC and apply_synapses no longer require mean_std.json. They still require mask.zarr.
• mean_std_{SGN,IHC,synapses}.template are renamed to mask_*. Only the job name changes. The pipelines, the README and the template comments use the new names.

Compatibility and merge order

• Merge this change before the cochlea-net change. With the old guards, the jobs of the new cochlea-net stop because mean_std.json is missing.
• The jobs also run with the current cochlea-net. Its first step still writes mean_std.json, and the prediction applies it.
• Archived jobs keep their mean_std_* names.

Follow-up

The mask_* jobs keep the resources of the mean/std step, for example 128G and 40 min for mask_SGN. Adapt them after the reportseff data of the first runs.