[cuda] fix half-size allocation of discretized gradient buffer #10240

Job	Run time
Python - oldest supported versions (manylinux_2_28, Python 3.9)	3m 22s
Python - latest versions (manylinux_2_28)	5m 21s
bdist wheel (manylinux2014-arm, gcc, Python 3.13)	2m 14s
mpi source (macos-15-intel, gcc, Python 3.12)	12m 1s
sdist (ubuntu22.04, clang-17, Python 3.13)	7m 59s
sdist (windows-2022, MSVC, Python 3.10)	8m 5s
bdist wheel (macos-14, clang, Python 3.11)	1m 12s
regular (ubuntu22.04, clang-17, Python 3.13)	8m 42s
sdist (macos-15-intel, clang, Python 3.9)	4m 22s
mpi wheel (ubuntu22.04, clang-17, Python 3.10)	9m 34s
mpi source (ubuntu22.04, clang-17, Python 3.13)	8m 37s
regular (macos-15-intel, clang, Python 3.11)	10m 30s
bdist wheel (windows-2022, MSVC, Python 3.13)	9m 23s
regular (manylinux_2_28, gcc, Python 3.11)	13m 18s
mpi source (manylinux_2_28, gcc, Python 3.10)	8m 2s
bdist wheel (manylinux_2_28, gcc, Python 3.10)	13m 20s
regular (windows-2022, MSVC, Python 3.11)	9m 33s
bdist wheel (macos-15-intel, clang, Python 3.13)	3m 0s
sdist (manylinux_2_28, gcc, Python 3.9)	11m 25s
gpu source (manylinux_2_28, gcc, Python 3.13)	13m 34s
mpi pip (ubuntu22.04, clang-17, Python 3.12)	1h 0m 2s
all-python-package-jobs-successful	5s
NuGet package	0s
	3h 43m 41s

Provide feedback