[Codegen][ROCm] Drop assume alignments

[qedawkins]

This reportedly hurts the LLVM optimizer. We may want to find a better way to annotate alignments.

[benvanik]

If this lands it should be tagged with an upstream issue to fix that - a backend should not pessimize codegen by having more information about pointer alignment. This is also a hard thing to profile as in many cases dynamically shaped models will be hurt the worst (where offsets are all dynamic values) - our suite definitely doesn't cover that.

[qedawkins]

Correctness issues without these assumes is scary...

Failed
Expected near equality of these values. Contents does not match.
  lhs:
    2x2x2xf32=[[3 6][15 18]][[21 24][78 84]]
  rhs:
    2x2x2xf32=[[1 1][1 1]][[2 2][2 2]]

[benvanik]

Methinks that codegen target needs some improvement :)

[qedawkins]

methinks so too :(