Fix sequence_length option in CLIPTokenizer

[james77777778]

Fix #2018

There is a bug in CLIPTokenizer where specifying sequence_length causes it to incorrectly use tf.RaggedTensor.to_tensor for padding, instead of self.pad_token_id.
Additionally, we need to compute padding_mask manually rather than relying on tf.RaggedTensor.to_tensor for the same reason.

This PR resolves the issue.

[mattdangerw]

@james77777778

Sorry for the delay! Finally catching up post holidays.

I'm not sure we really want to support this usage. I'll comment on the bug.

[mattdangerw]

Commented on the issue. #2018 (comment)

I think we probably just want to remove the "packing" feature from the tokenizer to agree with other tokenizer. But we probably need to think about our task design for CLIP more generally. Tentatively closing, but can re open if we need!