ensuring one shard per worker for sharded zarr arrays.

[yuriyzubov]

Summary

Fixes a data-loss race condition in the local alignment pipeline's transform-application step when writing to a sharded zarr v3 output array.

distributed_apply_transform was being called with align_chunksize (the zarr chunk size) as its dask blocksize. For an unsharded output that's fine -- one worker per chunk-file, no overlap. But once sharding groups several chunks into one shard, a shard (not the chunk) becomes the real atomic write unit: writing part of a shard requires zarr to read the whole shard, merge in the new chunk, and rewrite the whole shard (ShardingCodec._encode_partial_single). With multiple chunk-sized dask blocks landing in the same shard and no synchronization between workers, one worker's write can silently overwrite another's already-written chunk.

validate_processing_block_size exists specifically to catch "block too small" configurations like this, but it only ever checked the output's chunk size, not its shard size, so this exact configuration passed validation unflagged.

Changes

• bigstream/distutils.py: validate_processing_block_size now also checks the output array's shard size (when sharded), not just its chunk size. Extracted the shared compare/log/raise logic into _check_storage_unit_size to avoid duplicating it between the chunk and shard checks.
• bigstream/tools/main_local_align_pipeline.py: _align_local_data now passes a shard-sized block to distributed_apply_transform (getattr(align, 'shards', None) or align_chunk_size) instead of the raw chunk-sized align_chunksize, so each worker owns a whole shard and no two workers can ever write into the same one.
• tests/test_distutils.py (new): unit tests for validate_processing_block_size covering the unsharded case, the unsafe chunk-sized-block-on-sharded-array case, and the safe shard-sized-block case.

Test plan

• tests/test_distutils.py::test_unsharded_block -- unsharded array, block == chunk size, no error (existing behavior preserved)
• tests/test_distutils.py::test_processing_block_is_chunksize -- sharded array, chunk-sized block, now correctly raises ValueError
• tests/test_distutils.py::test_processing_block_is_shardsize -- sharded array, shard-sized block, no error

[cgoina]

Similar change needs to be done for distributed_invert_displacement_vector_field - This code actually has not been used in a very long time because the pipeline only computes the deformation field