Fix GPU introspection constants for architecture 1070

[csadorf]

Summary

cpp/include/nvforest/detail/gpu_introspection.hpp does not classify architecture 1070 as supporting 1,024 resident threads per SM. It therefore falls through to:

MAX_THREADS_PER_SM = 2048

Impact

MIN_BLOCKS_PER_SM is derived from MAX_THREADS_PER_SM / MAX_THREADS_PER_BLOCK. For architecture 1070, this produces 8 instead of 4.

This value is used by infer_kernel in its __launch_bounds__ declaration. As a result, downstream cuML source builds on architecture 1070 require a local patch adding __CUDA_ARCH__ == 1070 to the 1,024-thread branch.

Expected behavior

Architecture 1070 uses:

• MAX_THREADS_PER_SM = 1024
• MIN_BLOCKS_PER_SM = 4

Relevant code

• cpp/include/nvforest/detail/gpu_introspection.hpp
• cpp/include/nvforest/detail/infer_kernel/gpu.cuh