Fix datetime rounding errors.

[djkirkham]

Since the netcdftime.datetime object now has microsecond precision, some tests are failing due to rounding errors.

This change fixes the errors by introducing a function num2date_to_nearest_second, which attempts to replicate the old behaviour of netcdftime.num2date (although it accepts a cf_units.Unit rather than unit string and calendar arguments.)

I say 'attempts' because when num2date is given a value in seconds representing a time with a half second, it will round it up or down arbitrarily, and it is apparently impossible to replicate this behaviour with the new version. The new function will always round up.

[marqh]

I think that this function may prove to be useful for users experiencing issues with netcdf4 times

I think that it is worth considering making it public and adapting the first few comment lines as the docstring.

I would likely leave the note that this behaviour... as a developer comment and not include it in teh docstring in this case

[marqh]

this looks like a sensible and well crafted change. It deals with a set of known issues within netcdf4-python and should make Iris less fragile with respect to versions of netcdf4

the open question for me is the public/private nature of the iris.util function

[djkirkham]

the open question for me is the public/private nature of the iris.util function

Perhaps it could be added to the cf_units api, possibly as an optional argument to Unit.num2date

[marqh]

Perhaps it could be added to the cf_units api, possibly as an optional argument to Unit.num2date

I like this apporach, I think that num <> date is already in scope for cf_units

@bjlittle @pelson would you support this approach?
Is it useful for @djkirkham to propose such a change to cf_units for us to see and him to test?

[djkirkham]

The change has been made to cf_units (SciTools/cf-units#66) so this PR is no longer needed.