[Llama3.2-11b-vision] Add support for text-only inference through generator api

[skhorasganiTT]

Ticket

tenstorrent/vllm#53

Problem description

• Llama3.2-11b-vision did not support text-only prompts

What's changed

• Added support for text-only prompts in prefill by propagating text_only_inference to the cross attention transformer to skip cross attention layers entirely (possible since prefill is done with 1 user at a time)
• Added support for text-only prompts in decode by enabling the use of the full_text_mask in the TtLlamaCrossAttentionTransformerBlock::forward and propagating the full_text_mask_expand_11SD input
• Added a test to simple_vision_demo.py for processing a batch with mixed text-only and text-image prompts
• Deleted unused forward function in llama_vision_model.py
• (unrelated to changes above) updated old decode pcc in test_llama_cross_attention_transformer_text.py

Checklist

• Post commit CI passes
• Blackhole Post commit (if applicable)
• Model regression CI testing passes (if applicable)
• Device performance regression CI testing passes (if applicable)
• (For models and ops writers) Full new models tests passes
• New/Existing tests provide coverage for changes

[skhorasganiTT]

T3K unit/freq/demo tests: https://github.com/tenstorrent/tt-metal/actions/runs/12959616354
Single-card demo tests: https://github.com/tenstorrent/tt-metal/actions/runs/12959618787
All post-commit tests: https://github.com/tenstorrent/tt-metal/actions/runs/12959625351