Switch GLES wrappers over to using Clang (and fix inline assembly).

Bug: http://b/18003438

The GLES wrapper implementations make direct calls to their underlying
function without specifying that they are retaining the original
argument register(s). This change makes the retention of these argument
registers more explicit, and then removes the LOCAL_CLANG override,
since Clang no longer optimizes the code into a bad state.

We also switch to "naked" functions, so that we control all stack layout
information. This prevents us from accidentally still corrupting
arguments on x86 (and for other functions with many parameters).

Change-Id: I37f2ef4a697373388950b41f0a292393c9fcbac7