intel/eu/validate: Don't compute bogus access masks

For VxH indirects, the computed access mask is just bogus and for SEND instructions, the computation we were doing doesn't really make sense because the instruction width doesn't matter. The SEND case causes problems for thread spawner terminate messages which we emit at the full SIMD width of the shader even though they only read one header register.

Found by running UBSan on the back-end compiler unit tests.