Improve performance of GeoStack and GeoSeries by sharing and caching dims

[rafaqz]

When lazily loading GeoSeries the dims of the child files are being loaded/calculated for every file. This should be done once for the first file and copied for the rest. The assumption of the series is that all dims are the same.

This may break when it comes to attached metadata being different - that may need some subtlety to handle correctly.

GeoStack is also not sharing dims between layers in all cases. This is complicated by the fact that not all layers have the same dimensions - but they could choose from a list of shared dimensions, as is done internally in netcdf files.

[rafaqz]

This could also involve inheriting from AbstractDimDataset - which would add whole stack function application and indexing.

However - it's a special case where N = Mixed(), as the layers can have different numbers of dims. It also means each layer needs to track the types of the dims it has so they can be retrieved from the shared dims tuple.