NIR: update texop to be able to do more lowering

Changes in this series:

• deduct number of offset and grad components from sampler type, it seems more appropriate than using the coord_components, and with this change, hardware that requires lod, bias, and compare values to be part of the register that also passes the coordinate can merge these value into the coordinate source on the NIR level and still pass validation.
• Add flag to tex instruction to indicate that a cube texture access was lowered to an array2d access. Since txd still requires three-component gradient input, we need to indicate this lowering for validation.
• add a new R600 specific alu-op for cube coordinate and face evaluation