use mean of min/max years as offset in calculation of datetime64 mean

[kmuehlbauer]

• Closes Taking the mean of a long-spanning np.datetime64 array produces the wrong value #10019
• Tests added
• User visible changes (including notable bug fixes) are documented in whats-new.rst

This calculates the mean out of min and max years and uses that as the offset, effectively moving the offset right in the middle of the available datetime64. This decreases the maximum value of the calculated timedelta64 values by half, such that overflow is prevented.

[kmuehlbauer]

Merged the wrong PR 😞 in the wrong repo. Before I revert this, would it make sense to check if it's good enough, @spencerkclark?

[spencerkclark]

Thanks @kmuehlbauer—no worries. This is pretty clever—with this approach I think you can still end up in the odd situation where computing the offset can lead to overflow, e.g. if the minimum and maximum year are both 1677 with nanosecond-resolution times, but the resulting timedeltas and mean remain correct (at least to the extent that they would be without overflow). I think I'm on board with keeping this change.