Improvements to Pandas-to-Iris bridge

[trexfeathers]

🚀 Pull Request

Description

Full overhaul of iris.pandas.as_cube(). Closes #3324.

• More Pandas-esque handling of DataFrames, with each column assumed to represent a specific variable (i.e. a single Series). (Current model converts a DataFrame into a 2-dimensional Cube, treating the columns as the second dimension).
• Can specify columns that correspond to any type of _DimensionalMetadata (DimCoord, CellMeasure etcetera).
• Can create n-dimensional Cubes - using the specified dimension coordinates to determine the dimensions. If no dimension coordinates: falls back on the existing Pandas index, creating a 1D Cube.
• Can create multiple Cubes - one from each column not specified as dimensional metadata. All Cubes share the same dimensional metadata that is generated from the specified columns.

This is a breaking improvement, but is targeting Iris v3.3 which cannot include breaking changes since it is a minor release. I am therefore proposing using iris.FUTURE to allow users to opt in.

To do:

• [ ] Tests
• [ ] Alignment with @hsteptoe's work in #4669
• [x] calendar handling like the original implementation does
• [ ] Docstring updates Including example of how to reshape a DataFrame to make all data 1-dimensional
• [ ] What's New entry
• [x] Series/DataFrame agnosticism
• [x] Memory efficiency
  • Are we getting NumPy views, or copying the arrays?
  • How does this handle Dask DataFrames?

---

Consult Iris pull request check list