Add model compression support for se_t type descriptor

[denghuilu]

With the help of @liangadam , we have added the model compression support for the se_t type descriptor. The main changes are:

• Reconstruct the DPTabulate class for the support the of se_t type;
• Change the op name from TabulateFusion to TabulateFusionSeA for se_a type;
• Add op TabulateFusionSeT for se_t type model compression, and put the corresponding CPU and GPU implementations in tabulate.cc and tabulate.cu;
• Add enable_model_compression method for se_t.py;
• Note that the default step of se_t is 10 times larger than se_a.

By using the default water example benchmark system, we found that the lammps output of the se_t compressed model is consistent with the original model. And with the embedding net size setting[25, 50, 100], [4, 8, 16], the compressed model gains speedup by a factor of 20(single V100 GPU) and 12(CPU) correspondingly.

TODO:

• Add rocm support;
• Add unit test for 'se_t' model compression;
• Add doc support for 'se_t' model compression;
• Fix the bug of accuracy when the last_layer_size of embedding net is equal or less than 8.

[codecov-commenter]

Codecov Report

Merging #1175 (016bf73) into devel (4bc3b27) will increase coverage by 0.08%.
The diff coverage is 78.00%.

@@            Coverage Diff             @@
##            devel    #1175      +/-   ##
==========================================
+ Coverage   75.95%   76.03%   +0.08%     
==========================================
  Files          91       91              
  Lines        7260     7358      +98     
==========================================
+ Hits         5514     5595      +81     
- Misses       1746     1763      +17     

Impacted Files	Coverage Δ
deepmd/entrypoints/compress.py	76.36% <ø> (-0.83%)	⬇️
deepmd/entrypoints/main.py	88.77% <ø> (ø)
deepmd/utils/argcheck.py	90.03% <ø> (-0.04%)	⬇️
deepmd/utils/convert.py	15.78% <19.04%> (-1.16%)	⬇️
deepmd/entrypoints/convert.py	22.22% <20.00%> (-6.35%)	⬇️
deepmd/descriptor/se_a.py	95.31% <50.00%> (-0.95%)	⬇️
deepmd/utils/tabulate.py	88.20% <85.07%> (+5.06%)	⬆️
deepmd/descriptor/se_t.py	98.19% <100.00%> (+0.14%)	⬆️
source/op/_tabulate_grad.py	100.00% <100.00%> (ø)

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 4bc3b27...016bf73. Read the comment docs.

[liangadam]

An error ‘incompatible model version’ occurred when using the produced model.pb to run lammps.

terminate called after throwing an instance of 'std::runtime_error'
what(): incompatable model: version 1.1 in graph, but version 1.0 supported

[denghuilu]

An error ‘incompatible model version’ occurred when using the produced model.pb to run lammps.

terminate called after throwing an instance of 'std::runtime_error' what(): incompatable model: version 1.1 in graph, but version 1.0 supported

You should manually compile the deepmd-kit c++ interface from scratch to match the 1.1 model version.