Fix gradient not averaged when parallel training.

[shishaochen]

The injection of gradient averaging OPs happens at horovod.tensorflow.DistributedOptimizer.compute_gradients.

Before this pull request, each worker trains its own data without synchronization except variables' broadcast at the beginning.

This BUG may explain no acceleration on loss convergence when parallel training.

[codecov-commenter]

Codecov Report

Merging #1104 (639ec28) into devel (a5bdd14) will decrease coverage by 0.00%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##            devel    #1104      +/-   ##
==========================================
- Coverage   75.72%   75.71%   -0.01%     
==========================================
  Files          88       88              
  Lines        6998     6997       -1     
==========================================
- Hits         5299     5298       -1     
  Misses       1699     1699              

Impacted Files	Coverage Δ
deepmd/train/trainer.py	72.83% <100.00%> (-0.07%)	⬇️

---

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 a5bdd14...639ec28. Read the comment docs.