onnxruntime

Fix CUDA tests for Ampere cards, and bump layernorm tests opset version (#14761)

### Description

Three main changes:

* `qOrdered*` tests fail 100% on Ampere+ cards with cublas error.
Disable them on these cards.
* Bump LayerNormalization tests to opset 17, to be consistent with the
ONNX specification. Mark TRT EP as a provider that does not do error
checking on faulty LayerNorm definitions
* Remove null tensor for optional `bias` input for LayerNorm. Optional
inputs should be omitted entirely.


### Motivation and Context

Streamlines testing for CUDA and TRT with later NVIDIA architectures.

Signed-off-by: Kevin Chen <kevinch@nvidia.com>