speed up compute_small_box

[MoseyQAQ]

Summary
In the original implementation of compute_small_box() in the NEP class, several new GPU_Vector objects are created using cudaMalloc and then destroyed by cudaFree, which introduces a significant computational cost:

GPUMD/src/force/nep.cu

Lines 1387 to 1391 in ef31657

	GPU_Vector<int> NN_radial(type.size());
	GPU_Vector<int> NL_radial(size_x12);
	GPU_Vector<int> NN_angular(type.size());
	GPU_Vector<int> NL_angular(size_x12);
	GPU_Vector<float> r12(size_x12 * 6);

Modification
I added a new member variable small_box_data to store the relative information at the beginning to avoid recreating these new GPU_Vectors:

https://github.com/MoseyQAQ/GPUMD/blob/c00ef6d62c4a33cc922d2779b12f22d8e64753e1/src/force/nep.cuh#L111-L117

Verifications

For a given model.xyz, and a given nep.txt, I performed the following calculations using the old and new implementations separately :

• predict energy, force and virial, and compare them
• perform minimization, and compare optimized position, and corresponding energy, force, virial.
• perform 1ps-NPT simulation (1000 steps), and compare the position, cell, energy, force, and virial along the 1000 configurations.
  All the tests mentioned above have been passed, yielding the same results. All original input files and output files are attached.

Small_Cell_PR_test_upload.zip
(Note: raw trajectory files for NPT simulations are deleted, since they are too huge to be uploaded in GitHub issue)

Speed Benchmark

[brucefan1983]

Very nice speedup!