Description:

#6376 introduced an optimization to the Tile kernels to process inputs where the net tiling effect is just multiple copies of the input buffer.

For example:
input shape = [1, 1, 256 * 50]
repeats = [1, 200, 1]
output shape = [1, 200, 256 * 50]

This worked well when there was no batching involved and the optimization didn't kick-in when batching was introduced.
As a slight extension, handle batching in this optimization.

For example:
input shape = [5, 1, 256 * 50]
repeats = [1, 200, 1]
output shape = [5, 200, 256 * 50]

In this case, we would copy each of the 5 sub-tensors in the batch 200 times.

Improves the perf of a 1PP model by ~30% (95 percentile) when batch size is 5.

Motivation and Context
Performance