[mlir][nvgpu] Align integer WGMMA type checks with i8 operands (#212215)

Integer WGMMA uses i8 operands with an i32 accumulator, but the NVGPU
verifier and K-shape selection still checked for i16.

Changes both checks to i8 and adds tests for rejected i16 inputs and the
existing i8 "not supported yet" limitation.

GitOrigin-RevId: ed17053cb1381a578ce3243a0712ffe5aaaad35d