fix bug of mixed precision training

[denghuilu]

#1469 should be fixed by this PR.

[codecov-commenter]

Codecov Report

Merging #1471 (22be0f5) into devel (82c787d) will increase coverage by 0.29%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##            devel    #1471      +/-   ##
==========================================
+ Coverage   75.68%   75.98%   +0.29%     
==========================================
  Files          92       92              
  Lines        7671     7673       +2     
==========================================
+ Hits         5806     5830      +24     
+ Misses       1865     1843      -22     

Impacted Files	Coverage Δ
deepmd/train/trainer.py	74.53% <100.00%> (+2.08%)	⬆️
deepmd/utils/network.py	92.63% <100.00%> (+9.83%)	⬆️
deepmd/descriptor/se_a.py	94.82% <0.00%> (+0.64%)	⬆️
deepmd/fit/ener.py	91.81% <0.00%> (+0.90%)	⬆️

---

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 82c787d...22be0f5. Read the comment docs.

[wanghan-iapcm]

Do wen have similar issues in other descriptors?

[denghuilu]

Do wen have similar issues in other descriptors?

No, only se_a type support the mixed precision training.