Remove pad_with_end_token argument in CLIPTokenizer.

[james77777778]

Related to #2018

I have verified that CLIPTokenizer should always use end_token_id as pad_token_id in both CLIP and SD3.
I’m not sure why I initially implemented pad_with_end_token arg. It might be related to how the internal SD3 code was implemented.

CLIPPreprocessor is still needed for both CLIP and SD3. When I implemented it, I didn’t have a clear idea about adding a task class for CLIP.
@mattdangerw @divyashreepathihalli WDYT?

The script to verify the outputs:

import keras
import numpy as np
import transformers

import keras_hub

text = "a cat sitting on the table"

tokenizer = keras_hub.models.Tokenizer.from_preset("clip_vit_base_patch32")
preprocessor = keras_hub.models.CLIPPreprocessor(tokenizer, sequence_length=16)
keras_results = preprocessor(text)
print(keras_results)

tokenizer = transformers.CLIPTokenizerFast.from_pretrained(
    "openai/clip-vit-base-patch32"
)
transformers_results = tokenizer(text, padding="max_length", max_length=16)
print(transformers_results)

np.testing.assert_allclose(
    keras.ops.convert_to_numpy(keras_results["token_ids"]),
    transformers_results["input_ids"],
)
np.testing.assert_allclose(
    keras.ops.convert_to_numpy(keras_results["padding_mask"]),
    transformers_results["attention_mask"],
)