fix(arena): pass full conversation history to moderate_input in vision anony arena

[Chessing234]

Bug: In gradio_block_arena_vision_anony.py, moderate_input is called with text as both the user message and the all_conv_text argument:

text, image_flagged, csam_flag = moderate_input(
    state0, text, text, model_list, images, ip
)

The function signature is moderate_input(state, text, all_conv_text, model_list, images, ip), where all_conv_text is the full conversation history that the moderation filter checks. Passing only the current user message means prior model responses are never moderated, allowing policy-violating content to slip through multi-turn conversations.

Root cause: The other arena files (gradio_block_arena_vision_named.py) already compute all_conv_text from both models' prompts and pass it correctly. This file was left with the original two-text call.

Fix: Compute all_conv_text from both models' conversation histories (matching the pattern in gradio_block_arena_vision_named.py) and pass it as the third argument:

all_conv_text_left = states[0].conv.get_prompt()
all_conv_text_right = states[1].conv.get_prompt()
all_conv_text = (
    all_conv_text_left[-1000:] + all_conv_text_right[-1000:] + "\nuser: " + text
)
text, image_flagged, csam_flag = moderate_input(
    state0, text, all_conv_text, model_list, images, ip
)