Fix dataset map multiprocessing and PIL image byte loading

[Vedant-K1]

I was running the fine-tuning script and hit a couple of roadblocks... specifically running completely out of system RAM, and the inference test crashing on certain datasets. I've pushed a few fixes for this.

1. RAM OOM during dataset preparation
   When running map() on the datasets, it was hoarding memory and eventually crashing my setup.
   Fix: I added num_proc to speed things up with multiprocessing, but more importantly, I added writer_batch_size=75. This forces the datasets library to flush to disk regularly, which keeps the system RAM usage under control.

2. Inference test crashing on raw byte images
   If a HuggingFace dataset returns images as raw byte dicts instead of PIL Images, the pre/post-training inference test fails because the processor expects a PIL Image.
   Fix: Added a quick check to convert raw byte streams into RGB PIL Images using io.BytesIO.