Change the smooth method in se_atten descriptor

[iProzd]

This PR edits the smooth method of se_atten descriptor when using stripped type embedding:

1. Change the structure in stripped type embedding and compression ops.
2. Add smooth factor on the network output of type embedding.
3. Add a new descriptor se_atten_v2 to include these changes while keeping se_atten in old behaviors.

[iProzd]

The default values of stripped_type_embedding, smooth_type_embdding and set_davg_zero should be discussed.

[codecov]

Codecov Report

Patch coverage: 92.85% and no project coverage change.

Comparison is base (54cb07b) 78.73% compared to head (b31689e) 78.74%.
Report is 10 commits behind head on devel.

Additional details and impacted files

@@           Coverage Diff           @@
##            devel    #2755   +/-   ##
=======================================
  Coverage   78.73%   78.74%           
=======================================
  Files         241      242    +1     
  Lines       25582    25641   +59     
  Branches     1517     1517           
=======================================
+ Hits        20143    20191   +48     
- Misses       5049     5058    +9     
- Partials      390      392    +2     

Files Changed	Coverage Δ
deepmd/utils/finetune.py	85.36% <ø> (ø)
source/lib/src/tabulate.cc	62.46% <0.00%> (ø)
deepmd/descriptor/__init__.py	100.00% <100.00%> (ø)
deepmd/descriptor/se_atten.py	93.97% <100.00%> (+0.18%)	⬆️
deepmd/descriptor/se_atten_v2.py	100.00% <100.00%> (ø)
deepmd/entrypoints/train.py	88.00% <100.00%> (ø)
deepmd/utils/argcheck.py	95.95% <100.00%> (+0.06%)	⬆️

... and 5 files with indirect coverage changes

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