Repeated reading of attributes is harming performance

[davidhassell]

Repeated reading of attributes is harming performance - this is because every time h5netcdf is going back to self._h5ds to get the attributes, and this creates a new Variable every time. So even if a hd55/pyfive variable itself can cache its attributes, this is lost to h5netcdf.

This is also the case for dimensions, which are (often) associated with a variable (i.e. dimension coordinate variable), and so get their dimensions read.

The answer has to be (?) to apply store caching on the objects which are persistent in h5netcdf, such as Dimension and File.

[davidhassell]

I have a hacky branch (https://github.com/NCAS-CMS/h5netcdf/tree/dimension-access) where I'm playing with this