Describe the bug
The FoldPrePostReshapeSoftmax is added for LLaMA，but the conditions of FoldPrePostReshapeSoftmax are not strict enough, it will cause damage to other scenes. PR #1121 deletes FoldPrePostReshapeSoftmax in the general Transform, which will cause LLaMA matching errors.

To Reproduce
the input of the first reshape is inconsistent with the output of the second reshape.

Expected behavior
Write this reshape into LLaMA's pattern matching instead of as a general transform.

Origin model and code
p.Add<Passes.Rules.Neutral.FoldPrePostReshapeSoftmax>();
This line has been deleted, and it should be written into pattern matching unique to LLaMA

Environment (please complete the following information):

• OS: [Ubuntu20.04]
• nncase version [commit id 3817072]
• DL Framework [TensorFlow 2.5.0]

Additional context
Add any other context about the problem here.