Enhance classification task's model performance

[vinnamkim]

Summary

1. Memory caching

   • Fix the bug that the memory cache handler tries to resize and cache an image although there is no space on its memory pool.
   • Fix the bug that multiprocessing.Manager can be shutdown not in the main thread (dataloader's worker thread).
   • Reuse the resized image during caching

2. Change image decoder from OpenCV to PIL

   • Change the default image color channel format as RGB and remove the BGR to RGB color channel format conversion in the data preprocessor.

3. Set pin_memory=True in DataLoader.

4. Utilize a stacked tensor for a batch of images rather than a list of individual images.

   • This approach is necessary to eliminate the need for stacking image tensors in the main thread.

---

• This is the performance comparison according to this PR
  

---

How to test

I updated our unit tests related to this change and the integration tests must guard the correct training pipeline.

Checklist

• I have added unit tests to cover my changes.​
• I have added integration tests to cover my changes.​
• I have added e2e tests for validation.
• I have added the description of my changes into CHANGELOG in my target branch (e.g., CHANGELOG in develop).​
• I have updated the documentation in my target branch accordingly (e.g., documentation in develop).
• I have linked related issues.

License

• I submit my code changes under the same Apache License that covers the project.
  Feel free to contact the maintainers if that's a concern.
• I have updated the license header for each file (see an example below).

# Copyright (C) 2023 Intel Corporation
# SPDX-License-Identifier: Apache-2.0

[harimkang]

LGTM

[vinnamkim]

Converted to draft to fix https://github.com/openvinotoolkit/training_extensions/actions/runs/7384134398/attempts/1?pr=2749