Fix float128 on 32 bits

[franchuterivera]

This PR address the problem from #1127 .

Nevertheless, it is complicated as how many bits are used depends on the OS. For example:

In spite of the names, np.float96 and np.float128 provide only as much precision as np.longdouble, that is, 80 bits on most x86 machines and 64 bits in standard Windows builds.

This PR, at least will prevent auto-sklearn from crashing on this situations.

[codecov]

Codecov Report

Merging #1154 (c0cca43) into development (0982410) will decrease coverage by 0.03%.
The diff coverage is 0.00%.

@@               Coverage Diff               @@
##           development    #1154      +/-   ##
===============================================
- Coverage        85.83%   85.79%   -0.04%     
===============================================
  Files              137      137              
  Lines            10625    10625              
===============================================
- Hits              9120     9116       -4     
- Misses            1505     1509       +4     

Impacted Files	Coverage Δ
autosklearn/automl.py	84.75% <0.00%> (ø)
...eline/components/feature_preprocessing/fast_ica.py	91.30% <0.00%> (-6.53%)	⬇️
...ine/components/feature_preprocessing/kernel_pca.py	94.73% <0.00%> (-1.76%)	⬇️
autosklearn/ensemble_builder.py	76.65% <0.00%> (-0.41%)	⬇️
...ine/components/classification/gradient_boosting.py	92.17% <0.00%> (+0.86%)	⬆️
...ature_preprocessing/select_rates_classification.py	87.32% <0.00%> (+1.40%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 0982410...c0cca43. Read the comment docs.

[mfeurer]

This looks like a reasonable fix. As there is a specific np.float96, should we actively support that?