Make compatible with new keras Callback model

[ali-tny]

see https://keras.io/guides/writing_your_own_callbacks/ for guide - methods have been split to provide a different one for train/eval/predict time.

This change translates on_batch_end to on_train_batch_end rather than also for validation batches, so batch history only includes train batch loss - not sure if you wanna include the validation batches in the history here for plotting? Would just need to set on_test_batch_end to the same as on_train_batch_end.

[codecov]

Codecov Report

Merging #5 into master will increase coverage by 31.77%.
The diff coverage is 54.54%.

@@             Coverage Diff             @@
##           master       #5       +/-   ##
===========================================
+ Coverage   22.77%   54.54%   +31.77%     
===========================================
  Files           3        3               
  Lines         101      121       +20     
===========================================
+ Hits           23       66       +43     
+ Misses         78       55       -23     

Impacted Files	Coverage Δ
kerashistoryplot/callbacks.py	59.67% <54.54%> (+59.67%)	⬆️
kerashistoryplot/plot.py	17.14% <0.00%> (+17.14%)	⬆️

---

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 9a13745...5d9eb03. Read the comment docs.

[ig248]

Could you indicate what versions of keras the old vs new api would be compatible to? Also, maybe update the Callback docstring - tbh the only reason the base class is here is to avoid importing it from keras, hence avoiding keras as a dependency. However, if there are API differences, we could specify a keras version range in the setup.py?

[ali-tny]

I've updated to be backwards compatible (much like tensorflow.keras does). I considered just adding the keras dependency and importing the Callback base class, but to install tf.keras it needs all of tensorflow, and it seems kinda heavyweight now that its compatible across versions.

I also added a test - it requires adding matplotlib & ipython as dev dependencies though so it can import kerasplothistory.callbacks