stride units in numpy::from_data

Answering https://github.com/boostorg/python/issues/142, I notice that the `np::from_data` functions require strides to be counted in units of bytes, rather than elements. Is there any particular reason for that choice ? Given that the function also takes a `dtype` argument, which provides a `get_itemsize()` method, it seems plausible (and a bit more convenient) for callers to not have to use `sizeof(value_type)` explicitly.

@TallJimbo, any thoughts ?