[TF Led] Fix flaky TF Led test

[patrickvonplaten]

What does this PR do?

The reason why the TF LED test is flaky was not fully fixed in: #9459
and is actually the following:

Currently the decoder_attention_mask can have a 0 at its first input:

decoder_attention_mask[:, 0] == 0

Since the decoder uses a causal mask, this however leads to problems as a softmax over only very large negative numbers in computed. Now since TF and PT use slightly different large numbers, we can see significant differences between the models. The solution is to make sure that the decoder_attention_mask used for the tf_pt_equivalence test cannot be zero at the first position (I've done the same changes for all TFBart models in: #9497 and also made sure in #9497 that the TF templates are correctly updated )

[patrickvonplaten]

cc @LysandreJik @sgugger @jplu

[sgugger]

Thanks for fixing!