Add ViT attention plugin FMHA cubins and packed cu_seqlens support

[micwill755]

Summary

This PR adds generated ViT FMHA cubins from TensorRT-LLM into TensorRT-Edge-LLM under vitAttentionKernels and wires them into ViTAttentionPlugin. The plugin now supports both dense additive masks and packed cu_seqlens attention regions, enabling Qwen-style windowed ViT attention without requiring dense [S, S] masks.

Changes

• Added embedded ViT FMHA cubin assets and vit_fmha_cubin.h.
• Added ViT FMHA runner support for packed cu_seqlens.
• Added mask_type and max_seq_len plugin fields.
• Added support for Qwen2.5-VL head_dim=80.
• Added ONNX export support for trt::ViTAttentionPlugin.
• Updated Qwen2.5-VL visual ONNX export to emit ViT attention plugin nodes.
• Updated plugin format checks for mixed inputs:
• QKV/cos/sin/output: FP16
  • cu_seqlens: INT32
  • Dense additive masks: same dtype as QKV
• Validated end-to-end visual paths for Qwen2.5-VL, Llama 3.2 Vision/Mllama, and GR00T/Eagle/SigLIP-style models.

Testing

• Built NvInfer_edgellm_plugin.
• Ran ViT attention plugin correctness example.
• Ran end-to-end ViT attention plugin benchmark.
• Confirmed output shapes match PyTorch references across tested visual models.
• Exported Qwen2.5-VL visual ONNX with trt::ViTAttentionPlugin nodes in the graph.
• Built and serialized a TensorRT visual engine from the exported ONNX graph.
• Deserialized the saved TensorRT engine and ran inference with trtexec.
• Ran semantic correctness check comparing the deserialized TensorRT engine against PyTorch

[nvluxiaoz]

packed cu_seqlens attention regions - I think this is something we already supported. @nvamberl can you review and confirm if this MR adds additional features for Edge-LLM? Thanks!