[QNN EP] Fix "Zero tensor size!" for RandomUniformLike / RandomNormalLike

[qti-niscmami]

Description

Register the ONNX x input as a QnnTensorWrapper (APP_WRITE when it
is a fused-subgraph input) in the RandomUniformLike and
RandomNormalLike op builders, without adding it to the QNN node's
input_names. The QNN op still consumes only the synthesized static
*_shape tensor (and optional *_ort_qnn_ep_seed).

Motivation and Context

QNN's RandomUniformLike / RandomNormalLike ops only take a static
shape tensor.

The ONNX x input is used purely for shape inference at
compile time. The op builders correctly bake x's shape into a static
input, but never registered x itself as a tensor wrapper. ORT's
runtime input plumbing iterates the fused EP subgraph's inputs (which
still include x) and looks each one up in the wrapper map; with no
wrapper for x, the resulting list was empty and SetupTensors failed
its tensor_count == 0 guard.