Tensorflow 2.5 breaks KerasSequentialConverter

[landerlini]

I observed that loading with TF2.5.0 a keras Sequential model stored using TF 2.3.1 breaks scikinC because it cannot identify the output shape of the last layer without initializing the model first (which sounds a bit bizarre, but a simple workaround may exist).

[landerlini]

pytest testing succeeds with TF versions 2.5.0 and 2.6.0.

I will let this issue open and try to reproduce the problem.