Runtime performance optimization for ML

[zkhxsz]

for positions build, there are

[zkhxsz]

use from_blob build

[zkhxsz]

for 30000 atoms, use from_blob build

[zkhxsz]

use thrust::copy,30000 atoms

[zkhxsz]

can't overcome,mace need double