Speed up netCDF4, h5netcdf backends

[dcherian]

Accessing .shape on a netCDF4 variable is ~20-40ms. This can add up for large numbers of variables, e.g.: #9058

We already request the shape when creating NetCDF4ArrayWrapper, so we can reuse that.

• Closes opening datasets slow compared to netCDF4 #9058
• User visible changes (including notable bug fixes) are documented in whats-new.rst

[dcherian]

Turns out the same stuff works for h5netcdf too.

[dcherian]

@Illviljan Benchmarks didn't change but I don't think we run I/O benchmarks in CI anyway.

[Illviljan]

Files with many variables are my favorites! We have a few that are related to this that doesn't look like they are skipped:

xarray/asv_bench/benchmarks/dataset_io.py

Line 513 in 447e5a3

class IOReadSingleFile(IOSingleNetCDF):

This one has too few variables I guess? That could change.

xarray/asv_bench/benchmarks/dataset_io.py

Line 527 in 447e5a3

class IOReadCustomEngine:

This one uses a custom (and numpy array) backend, so probably won't hit the backend specific codes I guess.