fix(inference): keep transfer attention dense for single samples

[trungtpham]

Summary

• align multi-control transfer with the standard two-way single-sample inference path
• omit varlen metadata during inference so SM120 can select cuDNN instead of incompatible NATTEN kernels
• retain varlen attention for grad-enabled execution and add focused routing coverage

Test plan

• Run Python syntax checks on the modified files
• Run focused attention unit tests in the project CUDA environment
• Run video transfer inference on RTX PRO 6000 with the imaginaire4 v12 image
• Confirm H20/H100/A100 transfer inference remains healthy