Speed up open_tree

[felixcremer]

We have now an implementation which can load the data and where we can plot it but opening a single Sentinel-2 dataset takes 60 seconds and has 769 lock conflicts.

These lock conflicts are coming from Rasters._dims on the zarr variables and we should find a way to reduce this and maybe construct the dimensions differently.

[rafaqz]

Rasters._dims is very internal and can break any time, I wasn't aware anyone was using it. We really do need to centralise all this load code

[felixcremer]

I also was not aware that I am going to use it. I am only using it to walk around the problems I had with Rasters File Arrays on this Zarr data.

Yes we should work on centralising and unifying the loading code.