Adding minor numerical stability fixes for Llava

[hsilva664]

What does this PR do?

This PR attempts to further improve PR #34502 by adding some numerical stability changes after analyzing what was discussed in issue #35270. As discussed there, the instability reported is the result of NaN appearing sometimes because of the padding token. Since the padding token has a causal mask applied to all inner products where it appears, all its corresponding pre-attention weights are minus infinity, therefore, its corresponding output on the self-attention layer simply ends up being an average of all of the values, instead of the minus infinity having its intended effect of making the softmax entry equal to zero. This dynamic does not seem to co-exist too well with the normalization present in Llama (i.e. LlamaRMSNorm) and this sometimes causes NaN. The proposed changes simply attempt to impede the NaN to propagate to the valid sequence. They simply need to not leak to the valid tokens, since the subsequent FC layers (i.e. LlamaMLP) operate on a per-batch example basis (as opposed to, say, batch normalization, where features of one example would interfere with other examples from the batch), there is no risk of leakage there. In my preliminary tests, they do not seem to add significant computational overhead. Two main changes were added:

1. Making the values relative to padding tokens equal to zero explicitly. In fact, whenever a token was masked out against the whole sequence (via causal masking), the code makes its value explicitly zero. If the value becomes NaN, but the causal masking says it should not "count", the masking behaviour prevails over the NaN.
2. Making the causal masks pre-softmax weights (i.e. attn_weights) be calculated via tensor.masked_fill_ instead of it being a tensor sum. The mask that was previously being summed only seem to either contain zero or minus infinity. The suggested behaviour again makes the masking behaviour have precedence over the NaN behaviour.

Who might be interested?

@zucchini-nlp @ArthurZucker @pspdada