Central functions to open and normalize gridded datasets

[forman]

Expected behavior

Implement and use a single function open_gridded_dataset to open gridded datasets read from single files are from multiple files concatenated along time dimension. open_gridded_dataset should

• detect (also recursive) file wildcards and expand file list
• detect remote file access
• detect an appropriate internal chunking based on a configurable strategy. For example
  • "spatial": optimal for spatial analysis and visualization, that is chunking in spatial dimension taken from external NetCDF/HDF chunking or GeoTIFF tiling
  • "time": optimal for time analysis analysis: chunking mostly along time dimension.
  • "cube": optimal for spatio-temporal analyses
• open the dataset
• perform dataset normalization

The latter should make optional use of another normalize function that can be configured to

• rename spatial 1D longitude and latitude coordinates so in the end we have lon and lat coordinate variables
• detect a 0-360 degree longitude range and fix it to -180 to +180 degrees by rearranging variable grids
• ensure variables have a dimension time and we have a time coordinate variable given that attributes time_coverage_start and time_coverage_end are present
• ensure a coordinate variable named time has datatype np.datetime64
• ensure global spatio-temporal CF attributes are set

Actual behavior

There are many places in Cate's code where xr.open_dataset() are made without proper parameterization, e.g. appropriate chunking set. This has a major impact on performance and and also data compatibility due to missing normalization.

This is also related to #634, #623.

Specifications

Cate master as of 2018-05-10