[Feature] Improve overall performance of the optimizers

[kozistr]

Problem (Why?)

There're some inefficiencies

Solution (What/How?)

• remove .data
• torch.no_grad() to all step() function
• support bfloat16 dtype (XLA compatibility)

Other changes (bug fixes, small refactors)

version up to 0.4.0

Notes

nope

[codecov-commenter]

Codecov Report

Merging #51 (658d3a9) into main (3cd5158) will increase coverage by 0.74%.
The diff coverage is 87.93%.

@@            Coverage Diff             @@
##             main      #51      +/-   ##
==========================================
+ Coverage   87.16%   87.91%   +0.74%     
==========================================
  Files          26       26              
  Lines        1792     1812      +20     
==========================================
+ Hits         1562     1593      +31     
+ Misses        230      219      -11     

Impacted Files	Coverage Δ
pytorch_optimizer/version.py	0.00% <0.00%> (ø)
pytorch_optimizer/adahessian.py	62.36% <10.00%> (-0.28%)	⬇️
pytorch_optimizer/lookahead.py	87.32% <77.77%> (+0.55%)	⬆️
pytorch_optimizer/adabelief.py	86.55% <78.57%> (+3.62%)	⬆️
pytorch_optimizer/radam.py	91.01% <82.14%> (ø)
pytorch_optimizer/madgrad.py	96.38% <85.71%> (+1.32%)	⬆️
pytorch_optimizer/diffrgrad.py	94.62% <88.88%> (-1.08%)	⬇️
pytorch_optimizer/adabound.py	91.25% <89.47%> (+1.25%)	⬆️
pytorch_optimizer/ralamb.py	87.61% <90.32%> (-0.58%)	⬇️
pytorch_optimizer/ranger.py	95.91% <93.54%> (-1.86%)	⬇️
... and 9 more

---

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 3cd5158...658d3a9. Read the comment docs.