Fixes UT convergence problem (issue discussed in #1191)

[MostafaDehghani]

Set the default for "use_memory_as_final_state" flag to False.
Removed the undefined/reduntant flag "use_memory_as_last_state".

NOTE: PR #1192 submitted by @rllin-fathom is basically addressing the same issue, but here I also added the condition of using memory as the final state only when the transition function is an LSTM (to avoid unintentional use of memory for other transition functions).

[crystal0913]

Why does this parameter use_memory_as_final_state cause convergence problems?

[MostafaDehghani]

tl;dr
It was just a bug and the convergence of basic UT should be independent of this parameter.

use_memory_as_final_state is used when we use an LSTM as the transition function (use_memory_as_final_state=True means we use the memory instead of the output of LSTM as the output of the final step). I, unfortunately, set it to True in another PR and it messed up the code for the basic universal transformer. Sorry for the inconveniences!

[afrozenator]

@MostafaDehghani - thanks for updating the issues! Merging this now.