new_dd_object's array logic always assumes the metadata is numpy

[rjzamora]

There is a subtle limitation in the new_dd_object logic used to produce a dask.array.Array collection (important logic is here).

import cudf
import dask.dataframe as dd

s = cudf.Series(range(10))
ds = dd.from_pandas(s, 1)

def func(x):
    return x.values

meta = func(s)
result = ds.map_partitions(func, meta=meta)

assert type(result._meta) == type(meta)  # Fails, because meta is always a numpy array
assert type(result.compute()) == type(meta)  # Passes