Add animated GIF QC for per-instance and coloc crops

[Copilot]

Adds two new QC rules that produce animated GIFs of 2D SPIM crops centred on each detected instance, showing all stain channels side-by-side with a hollow circle marker for the equivalent spherical diameter and per-frame annotation.

New rules (qc.smk)

• qc_stain_instance_gif — {stain} wildcard selects the instance type; reads aggregated regionprops parquet; produces one GIF per atlas seg × stain × seg method
• qc_coloc_instance_gif — same layout for colocalized pairs; reads coloc parquet; gated on do_coloc

Both rules emit two outputs:

• *instancerandom.gif / *colocinstancerandom.gif — frames in reproducible randomised order
• *instancesorted.gif / *colocinstancesorted.gif — frames sorted ascending by equivalent spherical radius

Both are wired into rule all_qc.

New script (qc_instance_gif.py)

Shared by both rules via snakemake.params.instance_type = "stain" | "coloc".

Key behaviours:

• Loads all stain channels from the OME-Zarr; crops a patch_size×patch_size 2D axial slice at each instance's subject-space position via ZarrNii.crop_centered
• Draws a hollow yellow circle with radius = (3·nvoxels / 4π)^⅓ voxels (avg radius for coloc pairs)
• Annotation line per frame: idx | pos(x,y,z) | atlas ROI | radius
• Atlas labels looked up at runtime with ZarrNiiAtlas.label_region_properties; failure is logged and non-fatal
• Global per-channel intensity normalisation uses the coarsest available pyramid level (progressive fallback: base+5 → base+3 → base+1 → base)
• Sub-samples deterministically to max_instances when the table is large

patch_size and max_instances are configurable via instance_gif_patch_size / instance_gif_max_instances in the workflow config (defaults 100 / 200).

compute_colocalization.py

Added pos_coloc_x/y/z — subject-space midpoint of the two objects — to the coloc parquet. Required for crop_centered calls in the GIF script (the existing template_coloc_x/y/z are in template space only).