Fix/multimodal remove image reencode

[radik878]

Why was this change needed?

clevr_cogen_a_train and clevr_complex were re-encoding the same image to base64 twice:

• first in get_next_item,
• then again in collect_trajectories.

This added unnecessary CPU/memory work and relied on self.current_image shared state, which is fragile under concurrent workers.

What changed?

• Removed image re-encoding from collect_trajectories in both envs.
• Switched trajectory/message building to use base64_image from the input item tuple directly.
• Kept behavior unchanged from the caller perspective, but eliminated redundant allocations and reduced risk of image/item mismatch.