Fix mel length computation in Qwen2-Audio

[anferico]

What does this PR do?

1. Updates the misleading docstring in Qwen2AudioEncoder.forward stating that "Qwen2Audio does not support masking of the input_features", while in fact it does make use of it. This is likely a mistake made while copying the docstring over from Whisper
2. Fixes a bug that causes the last "valid" (i.e. not masked) frame of the sequence coming out of the first two convolutional layers to be masked incorrectly when max_mel_seq_len is an odd number

Here's a more elaborate explanation for point 2:
max_seq_len represents the maximum sequence length after the first 2 convolutional layers, of which only the first one compresses the input sequences (due to stride=2). The compression applied by the first convolutional layer can be computed using this expression (cf. torch.nn.Conv1d):

output_length = (input_length + 2*padding - dilation * (kernel_size - 1) - 1) // stride + 1

which, given that:

• padding = 1
• dilation = 1
• kernel_size = 3
• stride = 2

can be simplified to:

output_length = (input_length - 1) // 2 + 1

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

@ylacombe @amyeroberts @faychu

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: qwen2_audio

[anferico]

@ylacombe @amyeroberts @faychu any news?

[Rocketknight1]

cc @eustlb @ebezzam