Fix how the output tensor is created in CUDA SpaceDepth ops

[hariharans29]

Description:

The SpaceDepth ops are just thin wrappers over Transpose in that they do some re-ordering of the input based on a "virtual input shape" of the regular NCHW image input. The output tensor in these ops were being created based on the "virtual otuput shape" of this Transpose operation and then finally reshaped into the NCHW shape before the operation exits. The execution frame has a check to ensure that the requested output shape of the output buffer of the op be the same as the "expected" shape if the op output is also a graph output and if it isn't it produces a warning -

onnxruntime/onnxruntime/core/framework/execution_frame.cc

Line 806 in 4d0214f

LOGS(session_state_.Logger(), WARNING) << "Expected shape from model of " << *expected_shape

. This change adjusts the op's output tensor creation logic so that such warnings don't kick in should the SpaceDepth ops be producing graph outputs.

The existing op unit tests should keep this change covered.

Motivation and Context
Fix jarring warning seen by user in #11274