Vl dataloading + training improvements

[alay2shah]

Collate fn rework:

• Assistant-only label masking: Replaced TRL's default labels[pad] = -100 with label builder that masks everything except assistant content + <|im_end|> to match the internal stack (old collate had masking/padding issues)
• Per-sample error handling: Bad images skip the sample with a warning instead of crashing the batch
• Ray dataset integration: Switched from pre-materialized datasets to ray.train.get_dataset_shard() + ray_dataset_to_hf() consistent w/ prev. PR
• VLM processor params: max_image_tokens and do_image_splitting now flow from config through to load_vlm_model()

Processor configuration

• load_vlm_model() now accepts max_image_tokens and do_image_splitting
• Forces BICUBIC resampling (lost sometimes when starting from a checkpoint), padding config, cache disabled
• Replaced hardcoded max_image_tokens=256 with configurable parameter

PIL Image loading

• .convert("RGB") on all paths (handles RGBA, grayscale, palette)
• PIL safety settings (MAX_IMAGE_PIXELS, LOAD_TRUNCATED_IMAGES)
• is_image_loadable() to prevent breaks -- again everything to match internal

Distrubuted Ray data filtering for images (much faster)

• New is_valid_vlm_sft() row filter that validates message structure + actually tries loading each image
• validate_vlm_sft_format() now fails on unloadable images instead of just warning
• normalize_columns now covers vlm_sft type

[Rouzbehat78]

Passing the processor to the VLMTrainer to save the tokenizer, processor json files on checkpoint saving