While this PR may solve the issue with numpy.datetime64, it does not go far enough to also fix incorrect conversions of strings to float, e.g. "1E10" becomes 1e+10. See issue #266
I still think this PR can go forward, I just wanted to mention it.

Update: I discovered disable_numparse=True which solves that issue ... never mind ...

I argue in #351 that, if you have a type that converts both to int/float and to str, and you explicitly prefer the str representation -- that you should either:

1. convert the types to the desired str representation before youtabulate, or
2. ensure that at least one other value in the column is a str, causing tabulate to prefer their str representation

Otherwise, the (much more common) situation -- where you tabulate types that are convertible to int/float -- would be greatly interfered with by a change such as this.

Ah; as you found disable_numparse, I recommend we close this issue!

This PR needs test cases to cover and document all new edge cases, and how they are now handled.