Standardizing datasets dtypes

[justin-yan]

This PR follows up on discussion in #1900 to have an explicit set of basic dtypes for datasets.

This moves away from str(pyarrow.DataType) as the method of choice for creating dtypes, favoring an explicit mapping to a list of supported Value dtypes.

I believe in practice this should be backward compatible, since anyone previously using Value() would only have been able to use dtypes that had an identically named pyarrow factory function, which are all explicitly supported here, with float32 and float64 acting as the official datasets dtypes, which resolves the tension between double being the pyarrow dtype and float64 being the pyarrow type factory function.

[justin-yan]

@lhoestq - apologies for the multiple PRs, my previous one (#1905) got mangled due to some merge conflicts that I had trouble resolving so I just cherry-picked my changes onto a fresh branch here.

[lhoestq]

Nice thank you !