Fix sentinel token IDs in data collator for Flax T5 pretraining script

[rahuln]

What does this PR do?

Modifies the sentinel token IDs used in the data collator for the Flax T5 pretraining script so that they go in decreasing order starting at len(tokenizer) - 1, which matches the original T5 code.

Fixes #14282

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?

[LysandreJik]

Hey @rahuln! I'm pinging @patrickvonplaten for review as you have worked with him until now, but please note that he's off until next week so he'll review your PR when he's back! Thanks for your understanding.

[patrickvonplaten]

Great! Thanks a lot for digging into this issue and fixing it

[thomasw21]

Is that line not completely unecessary since we have the line right after?

[patrickvonplaten]

Well the next line makes use of the just changed sentinel_ids parrameter no?

[thomasw21]

Sorry what I meant is those two lines should be summarizable to

sentinel_ids = np.where(start_indices != 0, (len(self.tokenizer) - np.cumsum(start_indices, axis=-1)), 0)

Ie what is 0 is kept at 0 and what's not 0 is given a non 0 value, which means that the next where operation uses the same segmentation and thus overrides the values.

cc @patil-suraj