characterize-psf with larger datasets

[talonchandler]

Fixes #34.

I found that max_pool3d was the memory bottleneck, requesting more than 141 GB of our memory for daxi volumes (overfilling an H200).

Here I'm working around it by computing max_pool3d on the CPU. For smaller datasets, I found the CPU to be quite fast.

Related aside (should not block this merge): After removing this GPU memory bottleneck, I found that the next bottleneck is the fitting routine, which stalls on some beads. The tqdm progress bar shows steady progress on many beads, but very slow progress on some (likely not-very-gaussian) beads. @ieivanov I suspect that I'm not doing a good job filtering beads, and I'll chat with you about your routine for picking peak-fitting parameters.

[ieivanov]

I converted this to a draft PR. GPU compute on max_pool3d is very useful on regular-size images during live PSF tuning at the microscope. Let's find a better way to switch between CPU and GPU compute when needed. Thanks for catching that PSF fitting is the next slow step. Thats currently the bottleneck during live tuning also - max_pool3d takes 1.5 - 2 seconds, while PSF fitting can take 5-10 seconds, but so far that hasn't been slow enough to bother me much.