Avoid the use of shared variable.

[trivialfis]

• Use register instead.
• Fix Python benchmark script with latest pandas and XGBoost.

I ran the Python benchmark script using RTX8000 with CPU excluded:

PR

model	test_rows	cpu_time(s)	cpu_std	gpu_time(s)	gpu_std	speedup
adult-small	10000	0.0		0.011279034591279925	0.0071617705732274824
adult-med	10000	0.0		0.11670323801226914	0.011867081750644454
adult-large	10000	0.0		9.429356226988602	0.047229702183655395
covtype-small	10000	0.0		0.01151795060141012	0.003561754665202217
covtype-med	10000	0.0		0.6641728987917304	0.0032803489983452214
covtype-large	10000	0.0		74.24905845939648	0.5906834845783823
cal_housing-small	10000	0.0		0.006276315008290112	0.002279411050170562
cal_housing-med	10000	0.0		0.10911295339465141	0.0012658547831405737
cal_housing-large	10000	0.0		23.429275520995724	0.08867194402118761
fashion_mnist-small	10000	0.0		0.16429676760453732	0.06816532400946669
fashion_mnist-med	10000	0.0		4.30540003181668	0.07100682231932201
fashion_mnist-large	10000	0.0		178.59231641321676	1.0623992933829536

main

model	test_rows	cpu_time(s)	cpu_std	gpu_time(s)	gpu_std	speedup
adult-small	10000	0.0		0.010922096599824727	0.006964535125901965
adult-med	10000	0.0		0.11171154719777406	0.012917956737291719
adult-large	10000	0.0		9.12171239000745	0.16260832207140813
covtype-small	10000	0.0		0.011427397408988326	0.003081750313730569
covtype-med	10000	0.0		0.6582016726024449	0.010074246049065491
covtype-large	10000	0.0		73.67104159318842	0.6321216129573327
cal_housing-small	10000	0.0		0.005807545990683139	0.00245267488361271
cal_housing-med	10000	0.0		0.10978057580068708	0.0016856217806583914
cal_housing-large	10000	0.0		23.66984584759921	0.10060247707348238
fashion_mnist-small	10000	0.0		0.1685630762251094	0.06951830276227733
fashion_mnist-med	10000	0.0		4.44043413939653	0.06073396821402969
fashion_mnist-large	10000	0.0		183.3073539715959	0.6199842390017297

[trivialfis]

@hcho3 Have you seen this error using the rapids test matrix generator:

  File "/opt/conda/lib/python3.11/site-packages/rapids_dependency_file_generator/_rapids_dependency_file_generator.py", line 421, in make_dependency_files
    raise ValueError(f"No matching matrix found in '{include}' for: {matrix_combo}")
ValueError: No matching matrix found in 'cuda_version' for: {'cuda': '12.5', 'arch': 'x86_64', 'py': '3.11'}

with rapidsai/shared-workflows/.github/workflows/conda-cpp-tests.yaml@branch-24.08?

[bdice]

The failure on this PR is due to the addition of CUDA 12.5 support. This repository was missed during the upgrade. rapidsai/build-planning#73

I went ahead and pushed the changes you need in 444005e.

[bdice]

I also pushed commit e04ea9b with the changes from ci/release/update-version.sh 24.10.00. This repo is updated pretty irregularly and we need to keep all the RAPIDS versions aligned (not just the CI workflow branches).

[bdice]

Approving packaging changes. Thanks for the fix @trivialfis. Hope it's alright with the team that I am approving my own fixes, nothing should be controversial here (just missed version updates) and should unblock CI.