improve GLES jumptables

in the common case this saves one instructions per jump
(which will help with the i-cache).

this change also gets rid of the "use slow tls" option,
which was useless. So at least now architectures that don't have
assembly bindings will perform much better.

Change-Id: I31be6c06ad2136b50ef3a1ac14682d7812ad40d2