Store GL_FLOAT FlatInputs as fp32, not fp16.

There are two primary reasons for this:

 - GL_FLOAT FlatInput is primarily used for tests, and even more importantly,
   mostly accuracy tests. ATI's drivers appear to round off fp32 -> fp16
   wrong (truncate instead of round), which breaks some of these tests.
 - In case someone _would_ use GL_FLOAT inputs, they'd probably be updated
   every frame anyway, so the fp32 -> fp16 conversion step (probably on CPU)
   will negate any performance benefits by fp16 sampling anyway.