#1909 kernels: process all groups in fused GU and LDS down launches

Descripción

Closes #1908.

## Summary

When expert groups outnumber launch block rows, mode 8's fused gate/up and modes 7/8's LDS down kernels skip later groups. Stride the group index by `gridDim.y` in both kernels.

## What changed

Add a model-free regression to `native_grouped_parity`: 11 groups, varied entry counts, scattered outputs, nonzero entry offsets, and block-row counts 1/3/8. Check mode 7's down pass and mode 8's complete pass across five gate/up formats and two down formats. No launch defaults or arithmetic changes.

## Validation

MSVC/CUDA 12.4 build for `sm60;sm75`. On RTX 2080 Ti and P100, all 120 new comparisons per GPU fail before the fix and pass afterward; existing default-path checks remain bitwise equal. Targeted CTest and full CUDA integration build pass.

Draft pending the affected gfx906 HIP build; no AMD runtime result claimed.