Teacache implemented into hunyuan video framepack transformer model

[AChowdhury1211]

What does this PR do?

• This PR adds TeaCache support to HunyuanVideoFramepackTransformer3DModel, improving inference latency.

TeaCache is opt-in and can be enabled with:

transformer = HunyuanVideoFramepackTransformer3DModel.from_pretrained(
    "lllyasviel/FramePackI2V_HY", torch_dtype=torch.bfloat16
)
transformer.initialize_teacache()

What’s changed

• Added TeaCache logic in the model’s forward method.
• Introduced initialize_teacache() method.
• Wrote an initialization test inside the existing test for Hunyuan FramePack.

Let me know the fixes required

[AChowdhury1211]

@sayakpaul @yiyixuxu