[Feature] Optimizers with LR multipliers

[Chriskuei]

This PR implements optimizers with learning rate multipliers, which can be used to apply different learning rates for different layers. For example:

optimizer = MultiSGD(multipliers={'dense_1': 0.8})

In this example, parameter multipliers indicates that layers whose name contain dense_1 have a learning rate multiplier of 0.8 with respect to global learning rate.

Related issue

728

[codecov-io]

Codecov Report

Merging #729 into 2.2-dev will decrease coverage by 0.15%. The diff coverage is 92.62%.

@@             Coverage Diff             @@
##           2.2-dev     #729      +/-   ##
===========================================
- Coverage    94.09%   93.93%   -0.16%     
===========================================
  Files           91       99       +8     
  Lines         3115     3481     +366     
===========================================
+ Hits          2931     3270     +339     
- Misses         184      211      +27

Impacted Files	Coverage Δ
matchzoo/optimizers/multi_optimizer.py	87.5% <87.5%> (ø)
matchzoo/optimizers/multi_adam.py	88.13% <88.13%> (ø)
matchzoo/optimizers/multi_sgd.py	91.11% <91.11%> (ø)
matchzoo/optimizers/multi_adagrad.py	93.18% <93.18%> (ø)
matchzoo/optimizers/multi_rmsprop.py	93.18% <93.18%> (ø)
matchzoo/optimizers/multi_adadelta.py	93.87% <93.87%> (ø)
matchzoo/optimizers/multi_adamax.py	94.23% <94.23%> (ø)
matchzoo/optimizers/multi_nadam.py	96.49% <96.49%> (ø)
... and 6 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 a7c07ff...abae72f. Read the comment docs.

[bwanglzu]

should be a pr to keras rather than matchzoo.