Create a thread_local cache for textLocale

Bug:22378829

toLanguageTag is significantly more expensive than previously
thought (note ULOC_FULLNAME_CAPACITY is 157) and weighs
in at around 40us. Given that this is called on every Paint
and that there are typically thousands of Paint objects
created this adds up very quickly.

Given that the locale is almost always Locale.getDefault(),
a very simple thread_local cache of size 1 fixes this trivially

Change-Id: I819e60cac7a4b27e9dd5538332c22ce5bbd0851c