BUG: allow to import keras from tensorflow

[glemaitre]

closes #526

[codecov]

Codecov Report

Merging #532 into master will decrease coverage by 0.09%.
The diff coverage is 78.26%.

@@            Coverage Diff            @@
##           master     #532     +/-   ##
=========================================
- Coverage   98.93%   98.83%   -0.1%     
=========================================
  Files          85       85             
  Lines        5329     5343     +14     
=========================================
+ Hits         5272     5281      +9     
- Misses         57       62      +5

Impacted Files	Coverage Δ
imblearn/keras/_generator.py	91.54% <78.26%> (-6.7%)	⬇️

---

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 c0aa81c...cb4008b. Read the comment docs.

[glemaitre]

@chkoar The problem is not straightforward here.

Keras did not add the duck-typing on the Sequence for a long time. So we can get an error if we don't import Keras from the package used by the user (Keras or Tensorflow). I don't really have the solution there I think.

I thought about a hack that could use globals() to check if Keras has been imported and use the same package. However, this is pretty brittle. Do you have any idea?