Speed up safe_strtod and safe_strtof functions by using double-conversion library (#12102)

* Add double-conversion library to third_party

Link: https://github.com/google/double-conversion
Version: v3.0.0
With patch (add case insensivity for special values): https://github.com/google/double-conversion/pull/47

* Strnlen function to str_util.h
Returns the length of the given null-terminated byte string 'str'.
Returns strsz if the null character was not found in the first
'strsz' bytes of 'str'.

* Use double-conversion lib for safe_strto{d,f} implementation

* Path for double-conversion library is no longer needed, just use new version of this lib

* Use more obvious arguments names for Strnlen function

* Fixed DoubleToBuffer function in case of [full_precision_needed]
use safe_strtod insetead of raw inplementation

* add double-conversion compilation in makefile builds
- created compilation script: compile_double_conversion.sh
- added running compile_double_conversion.sh in {build_all_android.sh, build_all_ios.sh, build_all_linux.sh}
- added linking libdouble-conversion.a to build in tensorflow/contrib/makefile/Makefile

* Build double-conversion directly inside tensorflow/contrib/makefile/Makefile

* Put double_conversion to the rest of targets including number.* as src

* Sort external dependencies and make Sanity Checks happy

* Add test cases with trailing and leading whitespace characters to //tensorflow/core:lib_strings_numbers_test

* Remove octal numbers support from safe_strtod and safe_strtof

* Add double-conversion library to the cmake build

* Prepend lib/ to cmake lib path as per mrry instructions

See https://www.github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/cmake/external/zlib.cmake#L24

* Add  to double-conversion win32 build output path: fix up windows CMake build