Use make_tensor_proto / make_ndarray_from_tensor from tensorflow

[cp2587]

Because we wanted to avoid a dependency to tensorflow, we made the choice to re-write the make_tensor_proto / make_ndarray function of tensorflow. Doing this is dangerous as the code as been simplified (so have some limitations / bugs compared to TF maintained one).

TODO:

• add tensorflow as dependency and measure the impact on RAM consumption
• if this is too heavy as we expect it to be, find a way to still use the same implementation of make_tensor_proto / make_ndarray function than the TF one