Fix to ensure older tensorflow versions still work

[ascillitoe]

Changing two type annotations for tf.keras.optimizers.legacy.Optimizer to forward references since tf.keras.optimizers.legacy doesn't exist in older (<2.9?) tensorflow versions. This resulted in the following error:

>>> from alibi_detect.saving import save_detector
AttributeError: module 'tensorflow.keras.optimizers' has no attribute 'legacy'

We could define new type aliases like OptimizersTF in alibi_detect/utils/_types.py, but it seems to be overly complicating things to me (we'd need three groups, one containing classes only, one types only, and one containing both...). We're also already using forwardref's for the tensorflow optimizers in other places...

P.s. have tested this with tensorflow == 2.8 and the error no longer occurs (when it does with master).

[codecov]

Codecov Report

Merging #858 (2a986d8) into master (ce24f3c) will decrease coverage by 0.07%.
The diff coverage is 100.00%.

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #858      +/-   ##
==========================================
- Coverage   81.98%   81.91%   -0.07%     
==========================================
  Files         159      159              
  Lines       10375    10375              
==========================================
- Hits         8506     8499       -7     
- Misses       1869     1876       +7     

Files	Coverage Δ
alibi_detect/saving/_tensorflow/loading.py	85.20% <100.00%> (ø)
alibi_detect/saving/_tensorflow/saving.py	82.16% <100.00%> (ø)

... and 1 file with indirect coverage changes

[jklaise]

Nice one!