[Model] Add T5Gemma2 model plugin integration

[akh64bit]

This PR implements the T5Gemma2 encoder-decoder model as an out-of-tree vLLM plugin, moving the implementation from the core vllm repository as suggested in vllm-project/vllm#32617.

Changes

• Migrated the model implementation into vllm_bart_plugin/t5gemma2.py.
• Fixed the q, k, and v tensor reshaping bugs in attention inputs (MMEncoderAttention and vLLM's Attention expect (num_tokens, num_heads, head_dim), not (1, num_tokens, hidden_size)).
• Ensured residual connections are correctly placed in the T5Gemma2DecoderLayer block.
• Ensured the RoPE (rotary_emb) is correctly applied during the forward pass.
• Implemented SupportsMultiModal interface mapping.
• Updated Attention imports to work with the latest vLLM versions.
• Registered the model within vllm_bart_plugin/__init__.py.

How to test locally

To run this plugin alongside vLLM, ensure you have the required specific transformers fork and vLLM installed, then install the plugin in editable mode:

# 1. Clone and install the custom transformers library locally
git clone https://github.com/akh64bit/transformers.git -b t5gemma2
cd transformers
pip install -e .

# 2. Install the plugin in editable mode
cd ../bart-plugin
pip install -e .

# 3. Test loading the model using standard vLLM offline inference
python -c "
from vllm import LLM
# Will auto-load from the bart-plugin registry
llm = LLM(model='google/t5gemma-2-270m-270m', trust_remote_code=True)
print('Model loaded successfully!')
"

[akh64bit]

I've added an example_t5gemma2_usage.py script to demonstrate how to use the model with the plugin. You can test it out by running python example_t5gemma2_usage.py after following the setup instructions in the PR description.

[Bullish-Design]

Awesome! I've been looking for a good way to experiment with T5gemma2. How do you find its performance in comparison with other models of similar size? Are there any issues/limitations/things to be aware of when using it with vLLM?

[NickLucche]

Thanks for contributing @akh64bit !
Will look to get this merged after v0.16 support, as I see some of the changes are related to moving imports around (to sync with upstream)