Fix dilation>1 in Conv operator (GPU & CPU)

[harz05]

Implements #32 .

Conv generated wrong code for dilation > 1 because dilation was counted twice: fAttrKernelShape is expanded to the dilated size in Initialize and then used again with fAttrDilations. Same is already covered and fix as=> root-project/root#22473 / root-project/root#22474.

CPU path: ports the root-project/root#22474 one-liner, resetting fAttrDilations to 1 after the weight reorder so the dense im2col does not re-apply dilation.

GPU path (alpaka): applies the analogous fix in the generated kernels:

• the im2col kernel decodes its rows over the effective (expanded) kernel extents instead of the raw width, so the columns line up with the dilated _f layout
• the im2col gather no longer re-applies dilation
• _f is zeroed before the weight-vec kernel when dilation > 1 (alpaka buffers are not zero-initialised), so the gaps in the dilated layout stay 0

Generated im2col for a 3x3 dilation-2 conv, before -> after:

decode k_rem / 3u is changed to k_rem / 5u
gather oh*1u + kh*2u is changed to oh*1u + kh
zero (none) is changed to alpaka::memset(queue, deviceBuf_x_f, 0)

The changes are no-ops at dilation 1, so existing Conv tests are unaffected.

Tested on Colab T4: ConvWithDilation passes and all existing Conv tests still pass. Reverting only the codegen fix makes ConvWithDilation fail, confirming the fix.

[sanjibansg]

This would be very useful. We can proceed with testing and merge once the merge conflicts are fixed.

[harz05]

hi, i've rebased the PR, pls review and verify