vXi64 build_vectors are converted to v2Xi32 during type legalization
on RV32. This causes the use of SEW=32 splat instead of SEW=64.
This patch adds a DAG combine to detect the bitcasted pattern
and turn it into a SEW=64 splat.
We could maybe fix this in type legalization, but I'm not sure
we want to lose all zeroes/ones vector knowledge that early since
it's used by negate and not idioms for example.