While my observation is that most other Python packages don't import all subpackages.

"import urllib" does not also import urllib.parse, so you can't do:

  import urllib
  urllib.parse.quote("A&W")

but instead must explicitly import urllib.parse.

scikit-learn supports essentially the same use cases as NumPy and it doesn't import all its subpackages:

  >>> import sklearn
  >>> sklearn.linear_model
  Traceback (most recent call last):
    File "<stdin>", line 1, in <module>
  AttributeError: module 'sklearn' has no attribute 'linear_model'
  >>> import sklearn.linear_model
  >>> sklearn.linear_model
  <module 'sklearn.linear_model' from '[...]sklearn/linear_model/__init__.py'>

Thus leading to my earlier comment:

> The NumPy developers think "practicality beats purity" is more important, and their main use-case is long-running programs where startup costs are slow, which I interpret as meaning their special cases is special enough to break the rules.