DEP: linalg: improve deprecation of private modules named without underscore prefix

[j-bowhay]

Reference issue

toward #18279

What does this implement/fix?

Currently some of the deprecation warnings for private modules named without underscore prefix are misleading. For example:
Before this pr:

In [2]: from scipy.linalg.basic import levinson
<ipython-input-2-6918a53fe48c>:1: DeprecationWarning: Please use `levinson` from the `scipy.linalg` namespace, the `scipy.linalg.basic` namespace is deprecated.
  from scipy.linalg.basic import levinson

In [3]: from scipy.linalg import levinson
---------------------------------------------------------------------------
ImportError                               Traceback (most recent call last)
Cell In[3], line 1
----> 1 from scipy.linalg import levinson

ImportError: cannot import name 'levinson' from 'scipy.linalg' (/home/jakeb/development/scipy/build-install/lib/python3.10/site-packages/scipy/linalg/__init__.py)

After:

In [5]: from scipy.linalg.basic import levinson
<ipython-input-5-6918a53fe48c>:1: DeprecationWarning: `scipy.linalg.basic.levinson` is deprecated along with the `scipy.linalg.basic` namespace. `scipy.linalg.basic.levinson` will be removed in SciPy 1.13.0, and the `scipy.linalg.basic` namespace will be removed in SciPy 2.0.0.

This also reduces code duplication and strengthens testing.

Additional information

[j-bowhay]

Wasn't sure about the best course of action here. Everything in the __all__ of this module doesn't appear in scipy.linalg. This means if _sub_module_deprecation was used then they would all be stated for removal from __all__ in 1.13.0 but then there would be nothing left in the module. However, the module isn't due to be removed until 2.0.0 hence why I just modified the message directly.

[ilayn]

I didn't remove it just to stay out of your deprecation efforts, but indeed it is empty. So we can remove it since det and lu was living in here. And they are already available under linalg namespace.

[h-vetinari]

I find it strange that we're keeping around an empty module, but the warning has already promised that'd we'd only remove it for 2.0.0, so 🤷

[ilayn]

I guess nobody anticipated that we would touch Fortran code. But I think that promise is void now.

[h-vetinari]

See #19072

[h-vetinari]

Thanks! :)

[h-vetinari]

I find it strange that we're keeping around an empty module, but the warning has already promised that'd we'd only remove it for 2.0.0, so 🤷

[j-bowhay]

Everyone happy if we hit the button?

[h-vetinari]

Actually, this now needs to take into account the changes from #19067, e.g. we need to remove the np.cast aliases:

E   ImportError: cannot import name 'cast' from 'numpy'

[j-bowhay]

updated

[h-vetinari]

Thanks @j-bowhay! :)