[python] pass params to _InnerPredictor in train and cv and verbose fix

[StrikerRUS]

Fixed #1157.

@guolinke Is it critical that all params are passed? It seems that unknown (for _InnerPredictor) keys in params are just silently ignored and it's OK.

[guolinke]

I think it is okay.

[StrikerRUS]

Updated verbosity param description according to the following code:
https://github.com/Microsoft/LightGBM/blob/7b6f80f350f41d522535d5f62ef9ad7a0cbb0f85/src/io/config.cpp#L182-L190

Treat silent argument as Fatal log level (verbose=-1) to prevent showing warnings with verbose=-1 in params and silent=True passed at the same time like in this use case. Also, do not overwrite verbose param (if already specified by user) by silent argument (treat silent the same way as in sklearn wrapper).

Create Dataset in refit method silently.