Fix dilated Conv2d/Conv3d output size calculation

[wei-core]

Summary

• compute Conv2d P/Q from the effective dilated filter extent, (filter - 1) * dilation + 1
• apply the same correction to Conv3d Z for both symmetric and asymmetric padding constructors
• add regression tests for dilated 2D/3D problems, asymmetric padding and stride, and 1x1 filters

Background

The existing formula subtracts filter * dilation. It agrees with the standard convolution output-size formula only when dilation is 1. For dilation greater than 1 it undersizes the output, and a 1x1 filter is incorrectly made smaller by increasing dilation.

Contribution roles

I led the contribution direction and acceptance criteria, reviewed and approved
the final design, and reviewed every changed line. I understand the change and
own the final design decision and the submission. AI assistance was used for
repository research, root-cause analysis, implementation preparation, and
automated validation execution.

Testing

• reproduced the bug on the unmodified main branch: 2D output 3x4 instead of 4x6, and 3D output 4x3x4 instead of 5x4x6
• manually compiled and ran test/unit/conv/problem_size.cpp with MSVC 19.44, CUDA 12.9 headers, CCCL headers, and GoogleTest v1.14.0
• all 5 regression tests passed

Fixes#3502