The single address space SBA tracking sequence used CS GPR0/GPR1 as
scratch. The Direct Submission relaxed-ordering scheduler reads its
indirect MI_BATCH_BUFFER_START target from GPR0 and keeps GPR0-GPR11
live across task dispatch, so clobbering them corrupted the scheduler's
jump target, causing a GPU page fault at address 0x0 and device loss.
Use GPR13/GPR14 instead (GPR15 is already reserved for the SBA base).

Related-to: GSD-12387

Signed-off-by: Brandon Yates <brandon.yates@intel.com>
Source: c2c5fbde700b0ff89f16812eb20e14363828cd23