brw: Use the common NIR lowering for fquantize.
This generates one extra instruction to set the rounding mode to RTE due to f2f16_rtne in the lowering. If the rounding mode should be undef, as was previously being done in the backend, then we should probably fix up the core lowering as well.