[TensorRT EP] Fix bug for DDS output handling for empty tensor

[chilo-ms]

When the DDS output is empty tensor (i.e. any of the dimension is 0), TRT EP won't perform either cudaMemcpyAsync() nor cuda::Impl_Cast(), to prevent accidentally overwriting other location that might belong to other tensors.

This PR also refactors the code to only allocate single bytes for all empty tensors.

#TODO: add unit tests to cover the DDS code paths or doing more testing with concurrent,sequential, threaded faster-rcnn using onnx_test_runner and verifying outputs

[jywu-msft]

pls fix lint issues

[jywu-msft]

i think we can experiment with just allocating single byte for all zero tensor cases (currently we allocate number of bytes for that datatype).
or i wonder if we can even pre-allocate single byte and use the same address everywhere? (probably need to ask Nvidia/TensorRT what the requirement is)
that can potentially simplify the code even more.

[chilo-ms]

i think we can experiment with just allocating single byte for all zero tensor cases (currently we allocate number of bytes for that datatype). or i wonder if we can even pre-allocate single byte and use the same address everywhere? (probably need to ask Nvidia/TensorRT what the requirement is) that can potentially simplify the code even more.

I did test with ORT allocating only one byte allocation for empty tensor with faster-rcnn model and it works good.
Also, per TRT doc

If an engine binding is an empty tensor, it still needs a non-null memory address, and different tensors should have different addresses. This is consistent with the C++ rule that every object has a unique address, for example, new float[0] returns a non-null pointer. If using a memory allocator that might return a null pointer for zero bytes, ask for at least one byte instead.

Different tensor should have different address but we can have TRT EP to only allocate one dummy byte for each empty tensor