More protection against INT32 datasets that (might) contain -2^31 values

[LTLA]

For floating-point datasets, these should be coerced to the float; I don't think they are at the moment.

For integer datasets, we should just throw an error, because it's just impossible to preserve integer-ness. We can't even fall back to floating point like in alabaster.base because we need to stick by the type promises of extract_*_array()... so, just give up.