[CostModel][X86] Add test coverage for 'concat subvector' style shuffles

Shows 2 major issues:
 - SSE should be free as it splits everything to 128-bit
 - Negative costs for 128 -> 512 concat shuffles

Author: RKSimon