This PR proposes an API change (or rather an additional signature) for Dataset.select.

Currently we support passing per-dimensions select specifications as keyword arguments. This is generally quite convenient because in most cases dimensions are valid identifiers so the keyword syntax, e.g. .select(x=(0, 10)) provides the shortest and most convenient syntax. However when the dimension name is not a valid identifier, e.g. it's a string digit or contains other non-valid identifiers you have to write it out using dictionary unpacking:

x = hv.Dimension('0')

curve = hv.Curve([], kdims=[x])

curve.select(**{'0': (1, 3)})

This is not even particularly contrived because when you construct an element from a pandas DataFrame with default column names this is what happens. While a little cumbersome this use case at least works.

However, we are currently in the process of creating a new data interface for anndata and because of the complex data model we have to create special dimension objects, which are not easily mapped onto simple string names. This is where the current select approach completely breaks down since keyword arguments must be string based we cannot perform a select operation on elements backed by an anndata dataset, e.g.:

x = hv.Dimension('0')

curve = hv.Curve([], kdims=[x])

curve.select(**{x: 1})

Will error because x is not a string. Therefore I propose we overload the selection_expr argument for .select making it possible to instead write select operations as:

x = hv.Dimension('0')

curve = hv.Curve([], kdims=[x])

curve.select({x: 1})

This is fully backward compatible since it would previously just error. Before I do any more work on this I'd love to hear feedback.