ra: contiguous register class specialization

The core register allocator in Mesa is expensive at startup time on most GPUs because of its general-case support for register class q computation, and expensive in memory (at least) at runtime due to the large conflicts masks is has to maintain for each register.

Most GPU driver usage, though, is for classes representing contiguous registers in some base register file, and for that we don't need the general-case conflicts masks or the general-case q computation. Make a fast path for drivers that only do contiguous registers, and use it several drivers.