ENH: add quantile support for datetime64

[holymonson]

@eric-wieser, Is this OK? No explicit rounding.

close #11620

[tylerjereddy]

It might be helpful to add some tests to this potential contribution as well--I see that you've started discussing input and expected output with Eric in the linked issue so that might be a starting point.

[holymonson]

@tylerjereddy, testcase added. Travis CI reports false fail, please rebuild.

[tylerjereddy]

Ok, I restarted the failing Python 3.7 job in Travis

[shoyer]

A couple of thoughts:

• Consider casting to timedelta64 directly (e.g., dates.view('timedelta64')) instead of adding/subtracting an offset. This should be equivalent but more efficient.
• Please add tests to make sure that NaT is handled properly (like NaN).

[holymonson]

New commit with using view and adding testcase.

[seberg]

This is currently supported. However, I do not think we have any tests, so it would be great to recover (and maybe add to those). It may be tested downstream but I am not certain.

There is also the question of rounding that was brought up here, I am not immediately sure if this is a problem with our linear interpolation formula and datetime/timedelta inputs; it might be nice to vet it.

The handling of NaT is still broken, I suspect the fix should be closer to what we do for NaNs gh-20376 open about this.