Fix Booster read/write locale dependency

[AlbertoEAF]

This request fixes the bug mentioned on the issue #2890.

In C++, the process is by default is initialized to use the minimal "C" locale, however, in Java, when using the SWIG wrappers, this default is broken and the current machine locale is used instead.

Since during model read/write we rely on many C/C++ functions, some of which are 100% dependant on the locale, such as std::stod(), overriding the locale to a sane default is required.
Instead of setting the locale globally across the full process, as we might want to still read/write datasets according to the locale, a class to manage the locale context LocaleContext was created.
We only activate/deactivate that context (to force the "C" locale) during model reads/writes, to ensure our model format is independent of any locale changes in the machine or calling processes.

Motivation for this:

• In Python, the locale is always "C" by default, still as seen in the comment Invoking LightGBM via JEP has distinct behaviour than invoking outside of JEP (simple python script) ninia/jep#205 (comment), it is possible to break the model by changing locales.
• In Java, by using the SWIG interface, there isn't even a safety net like in Python, of using the "C" locale by default.

To avoid such issues, the fix should be done on the core library, so no bindings can be affected by locale issues, whatever happens, and this way the model format is always consistent.

[AlbertoEAF]

Hello @guolinke and @chivee,

can you tell me if this looks good to you and if so, to approve/revise?

[guolinke]

LGTM

[guolinke]

@StrikerRUS any style issues?

[StrikerRUS]

@AlbertoEAF Please fix the following cpplint errors:

[AlbertoEAF]

Changed everything you requested @StrikerRUS :)

Anything else?

[StrikerRUS]

@AlbertoEAF Thank you very much for fixing all linting issues!

[AlbertoEAF]

Thanks for merging @StrikerRUS :)