Support for automatic multidimensional loading & GC

[RainerHeintzmann]

The reshaping of filename arrays currently sounds non-ideal, or did I miss a tool which would do this already automatically?
How about selecting filenames by a Printf.jl-like pattern definition myimage_z=%.3d_t=%.2d.tiff or even allowing for multiple "*" in the pattern match, each of them automatically adding a dimension. Together with (optional) ranges and/or strides this could allow to flexibly select sub-arrays of data.
I did started something like this independently just based on FileIO.load, but the lazy loading mechanism in here might be a much better implementation.

Another thing which may be nice, is to turn off the garbage collector (GC) by default when loading a whole series into memory. The current version of FileIO.load has pretty bad timing when loading thousands of ".tiff" images, but turning the GC off temporarily avoids these issues.

[timholy]

The problem with turning off the GC is, what if the total number of files is such that the machine memory is insufficient to load them all? Users can of course disable the GC manually if they like.

[RainerHeintzmann]

I see. But maybe an option to do so as an argument to the function would point users in the right direction to speed up their loading time?

[timholy]

You can't turn off GC selectively. You could hold to a reference to any previously-loaded data, though. For example, instead of a single buffer you could have one buffer per image that gets populated when you load it. I'd support a PR to add this.

But I suspect the better approach is to load individual images virtually using memory-mapping. See https://juliaio.github.io/MultifileArrays.jl/dev/#Performance-tips. You could use StackViews and https://tamasnagy.com/TiffImages.jl/stable/examples/mmap_lazyio/#Memory-mapping-and-lazy-loading for a stack of TIFFs, for example.

See some details and comparisons here: https://tamasnagy.com/TiffImages.jl/stable/examples/mmap_lazyio/#Mechanism,-format-support,-and-performance