…nit barrier

…g parsing

Also move datetime string functions to their own source file.

…string

Also add some tests using the pytz library.

…tions

Also clean up the tests to work with the stricter date unit vs time unit
boundary that has been added.

The events don't fit well within the datetime, causing many special
cases and complicating the datetime code. A better way to proceed,
which received approval from Travis with regard to the reason events
were introduced in the first place, is to do the following:

* Generalize the structured arrays so they work with ufuncs
* Either use NPY_UBYTE, or introduce a new integer modulo N type
* The structured dtype [('date', 'M8[D]'), ('event', 'u8')], then
  will operate identically to the design of events in datetime.

These other changes will have much use elsewhere as well.

The default is 'same_kind', which allows data loss, but prevents
crossing the date unit/time unit boundary. As a special case,
printing dates with a timezone specified requires 'unsafe' casting.

…eptions

…tion

Since the datetime adds a new flexible dtype (datetime with generic units),
updating all the places where flexible dtypes are adjusted was error-prone.
Thus, this has been moved to a single place.

At the same time, I've added sizes for the various number types, so
they don't produce size-one strings by default anymore

…nction