Add tf.nn.gelu alias for TF >= 2.4

[aaronmondal]

Fixes #2252

See discussion in the above mentioned issue and the related, dropped PR #2256

[bot-of-gabrieldemarmiesse]

@AakashKumarNain

You are owner of some files modified in this pull request.
Would you kindly review the changes whenever you have the time to?
Thank you very much.

[aaronmondal]

Mentioning @seanpmorgan .

[seanpmorgan]

Looks very good thanks! Mind just modifying the version check to be consistent with elsewhere in the repo:
https://github.com/tensorflow/addons/blob/master/tensorflow_addons/losses/tests/quantiles_test.py#L88

[WindQAQ]

We should also warn users that the default value of approximate is changed from True to False if tf.nn.gelu is being used. Like

warnings.warn("Default value of `approximate` is changed from `True` to `False`")

[seanpmorgan]

Thanks! Marking blocked since this should be tested with #2250 . It shouldn't be long before the release now. Also please add the Warning Tzu-Wei suggested and then should be good.

[aaronmondal]

Added a warning in the docstring as well.

[seanpmorgan]

Hmm very slight difference w/ the keras implementation:

#23 357.3 =================================== FAILURES ===================================
#23 357.3 ______________ [doctest] tensorflow_addons.activations.gelu.gelu _______________
#23 357.3 [gw0] linux -- Python 3.6.12 /usr/local/bin/python
#23 357.3 051     Note that `approximate` will default to `False` from TensorFlow version 2.4 onwards.
#23 357.3 052     Consider using `tf.nn.gelu` instead.
#23 357.3 053 
#23 357.3 054     Usage:
#23 357.3 055 
#23 357.3 056     >>> tfa.options.TF_ADDONS_PY_OPS = True
#23 357.3 057     >>> x = tf.constant([-1.0, 0.0, 1.0])
#23 357.3 058     >>> tfa.activations.gelu(x, approximate=False)
#23 357.3 059     <tf.Tensor: shape=(3,), dtype=float32, numpy=array([-0.15865529,  0.        ,  0.8413447 ], dtype=float32)>
#23 357.3 060     >>> tfa.activations.gelu(x, approximate=True)
#23 357.3 Expected:
#23 357.3     <tf.Tensor: shape=(3,), dtype=float32, numpy=array([-0.158808,  0.      ,  0.841192], dtype=float32)>
#23 357.3 Got:
#23 357.3     <tf.Tensor: shape=(3,), dtype=float32, numpy=array([-0.15880796,  0.        ,  0.841192  ], dtype=float32)>

Is there anyway to increase doc test tolerance? Perhaps we just get rid of this doctest since gelu is on its way out? @WindQAQ

[WindQAQ]

I think we can just change the value of doctest. Actually, old c++ impl also gets different values from pure python ops. There is slight difference in casting and constant precision between c++ impl/python in addons/python in tf.nn.gelu.

[seanpmorgan]

LGTM thanks again @aaronmondal !