Evaluate multiclass logloss on the CUDA device

[yananlong]

Summary

CUDA multi_logloss currently falls back to the CPU, which requires copying the multiclass score matrix back from the device at evaluation time. Add a CUDA reduction for softmax and one-vs-all objectives, including sample weights, and compare the reported metric with predictions in the CUDA test suite.

Validation

• git diff --check and Python syntax checks pass.
• An earlier combined HIP/gfx1100 branch passed seven CUDA/HIP Python regression tests. The new weighted and one-vs-all cases have not yet run on this isolated current-main branch, so this PR is a draft.
• A prior single paired timing run showed 61.355 s for the correctness baseline and 56.917 s for the branch containing both this metric and the tree-transfer candidate. That combined result does not attribute the change to this PR; no standalone speedup claim is made here.