support neighbor stat on GPUs

[njzjz]

Fix #2619.

The GPU implementation in this PR is usually faster than the CPU in one thread (i.e., not using the feature implemented in #1624). Still, it needs parallelism in the batch dimension, which is blocked by #2618, regarding building the neighbor list. The GPU utilization is less than 10% for the water system. It should be improved when #2618 makes progress.

[codecov]

Codecov Report

Attention: 23 lines in your changes are missing coverage. Please review.

Comparison is base (f256dff) 75.46% compared to head (0ce3376) 75.87%.
Report is 3 commits behind head on devel.

Additional details and impacted files

@@            Coverage Diff             @@
##            devel    #2897      +/-   ##
==========================================
+ Coverage   75.46%   75.87%   +0.41%     
==========================================
  Files         244      245       +1     
  Lines       24522    24929     +407     
  Branches     1580     1615      +35     
==========================================
+ Hits        18505    18916     +411     
+ Misses       5086     5049      -37     
- Partials      931      964      +33     

Files	Coverage Δ
deepmd/utils/neighbor_stat.py	96.49% <100.00%> (+1.75%)	⬆️
source/op/custom_op.h	100.00% <ø> (ø)
source/op/prod_env_mat_multi_device.cc	73.35% <100.00%> (+12.92%)	⬆️
source/op/neighbor_stat.cc	72.61% <78.09%> (+6.26%)	⬆️

... and 16 files with indirect coverage changes

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[njzjz]

Let's run Test CUDA after #2892 is merged.

[wanghan-iapcm]

I am not an expert on this part of the code. I will listen to Denghui's review.

[denghuilu]

Looks good to me! If we can implement batched parallelism for this function, we should see a significant improvement in performance.