fix: handle empty one-dimensional tensors in concatenation

[shoumikhin]

Problem

Joining an empty tensor with a larger tensor can force a model to run that operation in PyTorch instead of TensorRT. This happens when a model starts with an empty cache and adds values to it.

PyTorch accepts an empty tensor with shape (0,) alongside tensors with more dimensions. TensorRT requires matching numbers of dimensions.

Change

Skip these empty tensors when checking dimensions and building the TensorRT concatenation. Continue rejecting dimension mismatches between the remaining tensors.

An empty tensor can still affect the result's data type. Choose the output type using every input before removing empty ones. Otherwise, later arithmetic could overflow in float16 when ordinary PyTorch would use float32.

The validator's test inputs include the data type it needs.

Tests

Passed 10/10 focused tests: four validator tests and six tests that build and run TensorRT engines. They cover empty tensors in different positions, positive and negative axes, mixed input types, and a float64 constant with conversion to float32 enabled. The empty-input validator regression fails on the base branch.

After rebasing, the existing nonzero suite passed 17 cases and skipped its RTX-only case, checking the nearby upstream converter changes.

The focused rerun used Linux x86_64, Python 3.12, PyTorch 2.15 nightly, and TensorRT 11.3 with the Python runtime. Earlier native-runtime coverage used TensorRT 11.2; a native TensorRT 11.3 build was not tested.

Windows, aarch64, TensorRT-RTX, TensorRT 10.x, and concatenation dimensions that vary between calls were not tested.