FIX: Deprecation of numpy.typeDict

[NiMlr]

This PR addresses #92.

The proposed alternative numpy.sctypeDict seems to be discouraged in numpy/numpy#17585.

Based on our relatively shallow current usage of numpy.typeDict:

mdp-toolkit/mdp/utils/routines.py

Line 295 in 300d288

_UNSAFE_DTYPES = [numx.typeDict[d] for d in

I propose to replace numpy.typeDict[d] by the generic numpy.dtype(d).type as:

# excluded the invalid inputs; see below
>>> [numpy.typeDict[d]==numpy.dtype(d).type for d in ['float16', 'float128', 'complex256']]
# [True, True, True]

This change to a function (as opposed to dictionary) has the downside of having to exclude the check:

mdp-toolkit/mdp/utils/routines.py

Line 297 in 300d288

if d in numx.typeDict]

However, we learn that exactly 2 of the current 5 checks will return false in the most recent release of numpy as of 03/27/22:

>>> 'float96' in numpy.typeDict or 'complex192' in numpy.typeDict
# False
>>> 'float16' in numpy.typeDict and 'float128' in numpy.typeDict and 'complex256' in numpy.typeDict
# True
>>> numpy.__version__
# '1.22.3'

I excluded the invalid ones and dropped the check. I think it is likely this will not create problems down the line.

General note:
I am not sure the functionality of checking for unsafe types generically is still up-to-date. Especially as it is only run if one is not using scipy.linalg.eigh:

mdp-toolkit/mdp/utils/routines.py

Line 286 in 300d288

if (_safe and not mdp.config.has_symeig == 'scipy.linalg.eigh'

[Stewori]

Since this passes local tests and given that #95 requires it to pass, I will merge this now.