[FEATURE] Add Donut & Flava model

[zhtmike]

Relies on Mbart #1195

What does this PR do?

Fixes # (issue)
In MS2.6/2.7, when there is a tied weights scenario, the load_param_into_net API may produce unexpected results during the weight loading stage. To address this, we switched to using the load_state_dict API, which more closely aligns with PyTorch's behavior.

Here is an example that demonstrates the buggy result caused by using the load_param_into_net API with tied weights.

import mindspore as ms
import mindspore.mint as mint


class Model(ms.nn.Cell):
    def __init__(self):
        super().__init__()
        self.layer1 = mint.nn.Linear(64, 64, bias=False)
        self.layer2 = mint.nn.Linear(64, 64, bias=False)
        self.layer1.weight = self.layer2.weight

# way 1
model = Model()
print(dict(model.parameters_and_names()).keys())  # layer 1 weight existed, layer 2 weight missing
model.load_state_dict({"layer1.weight": mint.ones((64, 64))}, strict=False) # can be loaded
print(model.layer2.weight.value()) # weight is correct

# way 2
model = Model()
print(dict(model.parameters_and_names()).keys())  # layer 1 weight existed, layer 2 weight missing
ms.load_param_into_net(model, {"layer1.weight": ms.Parameter(mint.ones((64, 64)))}, strict_load=False)
print(model.layer2.weight.value()) # weight is NOT correct

Adds # (feature)
Add Donut & Flava model

Donut

>>> from datasets import load_dataset
>>> from mindone.transformers import AutoProcessor, AutoModelForVision2Seq

>>> processor = AutoProcessor.from_pretrained("naver-clova-ix/donut-base-finetuned-docvqa", revision="refs/pr/23")
>>> model = AutoModelForVision2Seq.from_pretrained("naver-clova-ix/donut-base-finetuned-docvqa", revision="refs/pr/23)

>>> dataset = load_dataset("hf-internal-testing/example-documents", split="test")
>>> image = dataset[0]["image"]
>>> question = "What time is the coffee break?"
>>> task_prompt = f"<s_docvqa><s_question>{question}</s_question><s_answer>"
>>> inputs = processor(image, task_prompt, return_tensors="ms")

>>> outputs = model.generate(
...     input_ids=inputs.input_ids,
...     pixel_values=inputs.pixel_values,
...     max_length=512
... )
>>> answer = processor.decode(outputs[0], skip_special_tokens=True)
>>> print(answer)
"What time is the coffee break? 11-14 to 11:39 a.m."

version	mode	model	precision	task	s/step	weight load(s)
ms2.7.0	pynative	VisionEncoderDecoderModel	fp32	VQA	0.02	13

Flava

>>> from PIL import Image
>>> import requests
>>> from mindone.transformers import AutoProcessor, FlavaModel

>>> model = FlavaModel.from_pretrained("facebook/flava-full", revision="refs/pr/6")
>>> processor = AutoProcessor.from_pretrained("facebook/flava-full", revision="refs/pr/6")

>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
>>> image = Image.open(requests.get(url, stream=True).raw)

>>> inputs = processor(text=["a photo of a cat"], images=image, return_tensors="ms", padding=True)

>>> outputs = model(**inputs)

>>> image_embeddings = outputs.image_embeddings
>>> text_embeddings = outputs.text_embeddings
>>> multimodal_embeddings = outputs.multimodal_embeddings

>>> outputs.image_embeddings.shape
(1, 197, 768)

>>> text_embeddings.shape
(1, 7, 768)

>>> multimodal_embeddings.shape
(1, 205, 768)

version	mode	model	precision	task	s/step)	weight load(s)
ms2.7.0	pynative	FlavaModel	fp32	FeatureExtraction	0.07	15

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?
• Did you make sure to update the documentation with your changes? E.g. record bug fixes or new features in What's New. Here are the
  documentation guidelines
• Did you build and run the code without any errors?
• Did you report the running environment (NPU type/MS version) and performance in the doc? (better record it for data loading, model inference, or training tasks)
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

@xxx

[SamitHuang]

why change strict to False?

[zhtmike]

In the case of tied weights, there may be extra or missing parameters in the Hugging Face transformer checkpoint. Using strict=True will raise an error, so we follow the same design as the Transformers repo by setting strict=False.

[zhtmike]

/gemini review

[gemini-code-assist]

Code Review

This pull request introduces support for the Donut and Flava models, a significant and valuable addition to the library. The change to use load_state_dict for handling tied weights is also a crucial bug fix. The implementation of the new models appears to be a faithful port from the Hugging Face transformers library. However, I've identified several issues, including two critical bugs in the FlavaForPreTraining model that affect the model's output when return_dict=False. Additionally, there are a few minor typos in docstrings and log messages that should be corrected to improve clarity and usability. Addressing these points will greatly enhance the quality of this contribution.