direct conversions for dt64 units

[jbrockmendel]

ATM almost all datetimestruct conversions in tslibs use PANDAS_FR_ns. The exceptions are two functions in tslibs.conversions that convert non-nano datetime64 to datetime64[ns]. This PR replaces those two calls with a more performant implementation (asvs on the way).

The perf improvement is available b/c the functions currently convert dt64[not-ns] --> datetimestruct and then convert datetimestruct-->dt64[ns]. But for most units the conversion is just a multiplication, so we can cut out the convert-to-dtstruct step.

Finally np_datetime.convert_datetime_to_dtstruct is not actually used, but is included for exposition. After this PR, ujson notwithstanding (#19486), the remaining src/datetime files could be simplified quite a bit

[jbrockmendel]

In [4]: arr = np.arange(1, 100000, dtype='datetime64[h]')
In [6]: func = pd._libs.tslibs.conversion.ensure_datetime64ns
In [7]: %timeit func(arr)

Master:

In [8]: %timeit func(arr)
100 loops, best of 3: 3.55 ms per loop

In [9]: %timeit func(arr)
100 loops, best of 3: 3.58 ms per loop

In [10]: %timeit func(arr)
100 loops, best of 3: 3.58 ms per loop

PR:

In [8]: %timeit func(arr)
100 loops, best of 3: 2.28 ms per loop

In [9]: %timeit func(arr)
100 loops, best of 3: 2.28 ms per loop

In [10]: %timeit func(arr)
100 loops, best of 3: 2.28 ms per loop

But the year is about 6x slower and the month currently has a bug (that only got exposed when profiling, so apparently there is no test) so closing for now.