Further improvements to datetime roundtripping

[spencerkclark]

I forget if there was a reason we did not do this before, but this applies an approach I took in xarray (pydata/xarray#4684) to improve the accuracy of roundtripping np.datetime64[ns] dates to cftime.datetime objects. The changes here would mean that we'd do our absolute best to retain an integer dtype when encoding times through date2num.

I've included a pretty challenging test that passes on my machine; however it may not be a valid test on 32-bit systems. Edit: I updated the PR for what I think should allow things to pass on Windows.

[spencerkclark]

I'm not sure why the tests take so long in CI (previous runs took 3+ hours!). The full test suite runs in about 25 seconds on my local machine.

[jswhit]

I have no idea why the tests are so slow all of a sudden - they run super fast on my machine too.

[jswhit]

I think it's the code coverage that's slowing down the CI tests. I just disabled it in the CI workflow in your branch.

[jswhit]

This is really nice @spencerkclark, now we are very close to microsecond roundtrip accuracy. I think all this needs is a Changelog entry.

[spencerkclark]

Ah, disabling the code coverage seemed to do it, thanks! I went ahead and added a Changelog entry. Let me know if you need anything else.