[tmva][sofie] fix SAME_UPPER/SAME_LOWER autopad padding formula

[harz05]

Changes:

The SAME_UPPER/SAME_LOWER autopad calculation in ROperator_Conv.hxx and ROperator_Pool.hxx was using kernel/2 as a fixed pad value, ignoring input size and stride entirely. This gave wrong padding for any non-trivial case.

Fixed to use the correct ONNX formula:

output_size = ceil(input_size / stride)
total_pad = max(0, (output_size - 1) * stride + kernel - input_size)

Also fixed the direction: SAME_UPPER should place the extra pad at the end, not the beginning (it was swapped earlier).

Also added regression tests for both ConvWithAutopadSameLower and ConvWithAutopadSameUpper (kernel=3, stride=1 and stride=2 cases). Tests use odd-kernel models where padding is symmetric, since the existing Im2col utility only supports symmetric padding. Even-kernel SAME padding (asymmetric begin/end split) hits a separate pre-existing limitation in Im2col and I'll be adding that as an seperate issue.

Checklist:

• tested changes locally
• updated the docs (if necessary)

This PR fixes #21737