Add TensorRT conversion support for diffusion unet and hovernet

[binliunls]

Is your feature request related to a problem? Please describe. MONAI supported the TensorRT model conversion since 1.2. However, when tried to export the diffusion unet and hovernet to the TensorRT engine-based torchscript, there are some issues that some features need to be added to MONAI core to fix.

1. For the diffusion unet, the forward function needs two inputs, a tensor and an extra time step. While the trt_export API only supports one input shape and one dynamic batchsize now, these two parameters should be extended to support the multiple inputs case.
2. For the hovernet, the output of the forward function is a dict, which is not supported very well by the torchscript and TensorRT conversion. Therefore, a optional list output with a predefined order would be a good solution. There is a fixed similar issue of the detection network.

Describe the solution you'd like

1. Support the multiple inputs case by extending the input_shape and dynamic_batchsize parameters from Sequence[int] to Sequence[int] | Sequence[Sequence[int]] type, where each element in these two parameters stands for a shape requirement and a batch size range of one input.
2. Add a list output support option parameter for the hovernet like this PR and update the MONAI model zoo bundle accordingly.

[thalatamsainath]

Hi, its been a while since the last update on this. Has this been abandoned?