Initialize function performance issues on HCS datasets

[DavidStirling]

Following on from #116, the scanning process appears to be taking an unreasonably long time when working with some HCS datasets.

As an example, the BBBC017 dataset runs through bioformats2raw reasonably quickly to create our source Zarr, but raw2ometiff conversion gets hung up during initialisation. Data is available here, just one of the .zip files (1 plate) is sufficient. This is all running with default arguments.

There are a few thousand series within the input zarr. From a quick look it seems like examining the contents of each series takes ~0.2s each to start, which isn't ideal. However by the time we're at s=2000 they're taking ~2.5 seconds each. Every image in the dataset should be of similar size/format so I can't see an obvious reason to expect this beyond a bug somewhere.