Add inference script and support for different input image shapes

[Edmund-a7]

Brief:

• Add H and W parameters in vggt.py、eval_utils.py、aggregator.py、attention.py，enabling the model to automatically handle image sizes.
• Add inference.py

Hi! I've made a couple of improvements to the project.

Problem:

1. The repository was missing an inference.py script to run the model on custom datasets.
2. In the attention mechanism, the token height and width were hardcoded. This isn't ideal, as they should adapt to the processed image dimensions.

Solution:

1. I've added a new inference.py script.
2. I've modified the attention code to automatically calculate the tokens based on the image size (specifically, W / 14 and H / 14 in attention.py:161).

I have tested the code successfully. This makes the model more robust and easier to use for inference. Please let me know if you have any feedback!

[mystorm16]

Hi @Edmund-a7, thank you for your work on FastVGGT. We’ve updated the evaluation for the custom dataset, including the related point cloud and pose visualizations. Following your suggestion, we also modified the hardcoded height and width. Please give it a try!