Now we use the CTAB data to figure out where the high-level code actually
 thinks the bounds of the array should be, and allocate just to that range
 when we see relative addressing.

This will let the GL glue only push a little data, instead of a wide range
 that contains both possible accesses, and also drop the effort to pepper in
 DEF constants to the pushed data.

I hope this will prove to be a big win, but we'll see.

This will now blow up if there are relative addressing in shaders without a
 CTAB block, but anything generated by the HLSL compiler should have it, so
 I hope this will be "good enough."

--HG--
branch : trunk