This is a followup to PR numpy#5505, which didn't go quite far
enough.  This fixes two issues in particular:

1) The record class also needs an updated `__getitem__` that works
analogously to its `__getattribute__` so that a nested record is
returned as a `record` object and not a plain `np.void`.  In other
words the old behavior is:

```python
>>> rec = np.rec.array([('abc ', (1,1), 1), ('abc', (2,3), 1)],
...       dtype=[('foo', 'S4'), ('bar', [('A', int), ('B', int)]),
('baz', int)])
>>> rec[0].bar
(1, 1)
>>> type(rec[0].bar)
<class 'numpy.record'>
>>> type(rec[0]['bar'])
<type 'numpy.void'>
```

demonstrated inconsistency between `.bar` and `['bar']` on the record
object.  The new behavior is:

```python
>>> type(rec[0]['bar'])
<class 'numpy.record'>
```

2) The second issue is more subtle.  The fix to numpy#5505 used the
`obj.dtype.descr` attribute to create a new dtype of type `record`.
However, this does not recreate the correct type if the fields are
not aligned.  To demonstrate:

```python
>>> dt = np.dtype({'C': ('S5', 0), 'D': ('S5', 6)})
>>> dt.fields
dict_proxy({'C': (dtype('S5'), 0), 'D': (dtype('S5'), 6)})
>>> dt.descr
[('C', '|S5'), ('', '|V1'), ('D', '|S5')]
>>> new_dt = np.dtype((np.record, dt.descr))
>>> new_dt
dtype((numpy.record, [('C', 'S5'), ('f1', 'V1'), ('D', 'S5')]))
>>> new_dt.fields
dict_proxy({'f1': (dtype('V1'), 5), 'C': (dtype('S5'), 0), 'D':
(dtype('S5'), 6)})
```

Using the `fields` dict to construct the new type reconstructs the
correct type with the correct offsets:

```python
>>> new_dt2 = np.dtype((np.record, dt.fields))
>>> new_dt2.fields
dict_proxy({'C': (dtype('S5'), 0), 'D': (dtype('S5'), 6)})
```

(Note: This is based on numpy#5920 for convenience, but I could decouple
the changes if that's preferable.)