[Tosa] : Equalize ranks for all operands for tosa.select + Slice conv inputs for dynamic batch as long as spatial dims are static.

[sahas3]

Fixes a bug where not all operands of tosa.select had same rank. Looked at other uses of tosa.select op in the code and they all looked correct to me.

[sahas3]

@justin-ngo-arm : just for clarification was it a conscious choice to not use CreateOpAndInferShape calls for creating tosa ops here? I see that command being used in the tfl-to-tosa lowering on TensorFlow side. Thanks!

[sahas3]

Fixes a bug where not all operands of tosa.select had same rank. Looked at other uses of tosa.select op in the code and they all looked correct to me.

Added another bugfix to handle dynamic batch when slicing inputs are required to honor the checks like 'tosa.conv2d' op expected input_height - 1 + pad_top + pad_bottom - (kernel_height - 1) * dilation_y to be wholly divisible by stride_y

[justin-ngo-arm]

@justin-ngo-arm : just for clarification was it a conscious choice to not use CreateOpAndInferShape calls for creating tosa ops here? I see that command being used in the tfl-to-tosa lowering on TensorFlow side. Thanks!

At one point in the past, I was working on refactoring to use CreateOpAndInferShape for TorchToTosa. However, TOSA is now more restrictive about broadcasting and same rank (with the SameOperandsAndResultsRank trait), which makes it more complicated to make that transition to CreateOpAndInferShape.

cc: @sjarus