[bug] fix for incorrect output from grouped Conv and regression test for Conv with batch>1 & group>1

[harz05]

Closes #23

In Generate_GPU_ALPAKA and GetBlasConfig of ROperator_Conv.hxx, gemm_n was being set to total output channels and used directly in per-group matmul calls and BLAS layout config without dividing by fAttrGroup for the grouped path. The comment mentions that "we divide per group at launch" but the division was missing and same was verified with test failure as reported by #22 ,this caused per-group matmuls to compute all output channels instead of just the group's slice.

Restoring the division at the variable assignment fixes it in one spot for both call sites.

EDIT: The PR nowadds a ConvGroupBatch GTest (group=2, batch=4) as the regression test for this path; it fails without this fix and passes with it.

[harz05]

@sanjibansg a quick check on this. I noticed that in the recent operator init refactor, the gemm_n /= fAttrGroup division got dropped, was that intentional or did it just not carry over from the old Initialize() setup? I've added it back here in both Generate_GPU_ALPAKA and GetBlasConfig. Happy to change if you had a different intent for handling the grouped path.

[harz05]

also added a ConvGroupBatch GTest that exercises group>1 and batch>1 together (group=2, batch=4), which was never covered. Verified on Colab T4 with the tests passing