[backend][amd] Reorder and adjust optimization passes

[antiagainst]

This commit adjusts optimization passes a bit:

• Invoke canonicalization and cse after initial conversion to LLVM dialect in MLIR. This makes it possible to pick up canonicalization in other dialects (e.g., scf) before fully going down to LLVM dialect, which has not much canonicalization by design, or LLVM proper.
• First attach kernel attributes then link in external libs and run general LLVM optimize. Those kernel attributes may affect how optimizations are done.

[jlebar]

This makes it possible to pick up canonicalization in other dialects (e.g., scf) before fully going down to LLVM dialect

What's an example of this?

[antiagainst]

This makes it possible to pick up canonicalization in other dialects (e.g., scf) before fully going down to LLVM dialect

What's an example of this?

When debugging print op lowering, I found the following after convert-triton-amdgpu-to-llvm:

    %46 = llvm.mlir.constant(true) : i1 loc(#loc)
    %47 = llvm.addrspacecast %45 : !llvm.ptr<1> to !llvm.ptr loc(#loc)
    %48 = scf.if %46 -> (i32) {
      %141 = llvm.load %47 : !llvm.ptr -> i32 loc(#loc)
      scf.yield %141 : i32 loc(#loc)
    } else {
      %141 = llvm.mlir.constant(0 : i32) : i32 loc(#loc)
      scf.yield %141 : i32 loc(#loc)
    } loc(#loc)

It carries all the way down to converting to llvm ir proper. not a major issue. but still good to have it cleaned a bit.