Add --reset_learning_rate option to lstmtraining

[nagadomi]

When the --reset_learning_rate option is specified, it resets the learning rate stored in each layer of the network loaded with --continue_from to the value specified by the --learning_rate option.
If checkpoint is available, it does nothing.

background:

Currently, the -learning_rate option specified when finetuning with the lstmtraining --continue_from ... command is ignored. Instead, the learning rate stored in the pretrained model will be used. You can check its value with the combine_tessdata -l command.
The learning rate stored in the pretrained model is different for each language, and the behavior of finetuning is also different.
So, if the user has experience in machine learning, it is useful in some situations to be able to manually override the learning rate.

# finetuning with a low learning rate
lstmtraining --learning_rate 0.0001 --reset_learning_rate --continue_from ...

# training from a high learning rate using the pretraiend model as initial parameters
lstmtraining --learning_rate 0.001 --reset_learning_rate --continue_from ...

# learning rate decay manually
lstmtraining --learning_rate 0.0003 --reset_learning_rate --model_output ./lr1 --continue_from ...
lstmtraining --learning_rate 0.0002 --reset_learning_rate --model_output ./lr2 --continue_from ./lr1_checkpoint ...
lstmtraining --learning_rate 0.0001 --reset_learning_rate --model_output ./lr3 --continue_from ./lr2_checkpoint ...
lstmtraining --learning_rate 0.00001 --reset_learning_rate --model_output ./final --continue_from ./lr3_checkpoint ...