Closes #4463

As raised in #4463 when regridding with iris.analysis.UnstructuredNearest the dtype wasn't being preserved and the mask was being thrown away. These issues turned out to be linked.

Regarding the dtype, previous changes to the underlying trajectory code chose to preserve the behaviour from the initial Iris commit (10 years ago!) that created a resulting new_cube with an empty data array, then filled it.

The other issue was the mask not being preserved. This seemed to be a consequence of filling the empty array, which currently looks like:

As shown in the example below, if I try to fill any empty array with a masked array, the mask is simply thrown away. This is why I suspect we were first filling with the mdi value

>>> source_data = np.ma.array([1, 4], mask=[0,1], dtype=np.float32)
>>> new_data = np.empty((1,2))
>>> new_data
array([[6.94805265e-310, 6.94805265e-310]])
>>> new_data[:] = source_data
>>> new_data
array([[1., 4.]])

But this won't happen if we just assign cube data to our new data array.

In this PR, in the first commit I have first added a couple unit tests to check the dtype and mask are preserved. The second commit include the fix which makes the tests pass.