Canonicalize the IR before optimization

[maleadt]

Specifically, convert every scalar SSA value into a Tile. This simplifies compilation, and matches cuTile Python.

Before:

julia> ct.code_structured(vadd, Tuple{ct.TileArray{Float32, 1, ct.ArraySpec{1}(128, true, (0,), (32,))},
                                 ct.TileArray{Float32, 1, ct.ArraySpec{1}(128, true, (0,), (32,))},
                                 ct.TileArray{Float32, 1, ct.ArraySpec{1}(128, true, (0,), (32,))},
                                 ct.Constant{Int64, 16}})
1-element Vector{Pair{IRStructurizer.StructuredIRCode, DataType}}:
 StructuredIRCode(
│ %20 = cuTile.MakeTokenNode()::cuTile.TokenType()
│ %1  = cuTile.Intrinsics.get_tile_block_id(0)::Int32
│ %3  = cuTile.Intrinsics.make_tensor_view(_2)::cuTile.TensorView{Float32, 1}
│ %4  = cuTile.Intrinsics.make_partition_view(%3, (16,), cuTile.PaddingMode.Undetermined, cuTile.nothing)::cuTile.PartitionView{Float32, 1, Tuple{16}}
│ %6  = Core.tuple(%1)::Tuple{Int32}
│ %7  = cuTile.Intrinsics.load_partition_view(%4, cuTile.nothing, cuTile.nothing, %6, %20)::cuTile.FloatTile{Tuple{16}, Float32}
│ %8  = cuTile.Intrinsics.make_tensor_view(_3)::cuTile.TensorView{Float32, 1}
│ %9  = cuTile.Intrinsics.make_partition_view(%8, (16,), cuTile.PaddingMode.Undetermined, cuTile.nothing)::cuTile.PartitionView{Float32, 1, Tuple{16}}
│ %11 = Core.tuple(%1)::Tuple{Int32}
│ %12 = cuTile.Intrinsics.load_partition_view(%9, cuTile.nothing, cuTile.nothing, %11, %20)::cuTile.FloatTile{Tuple{16}, Float32}
│ %13 = cuTile.Intrinsics.addf(%7, %12)::cuTile.FloatTile{Tuple{16}, Float32}
│ %15 = Core.tuple(%1)::Tuple{Int32}
│ %16 = cuTile.Intrinsics.make_tensor_view(_4)::cuTile.TensorView{Float32, 1}
│ %17 = cuTile.Intrinsics.make_partition_view(%16, (16,), cuTile.PaddingMode.Undetermined, nothing)::cuTile.PartitionView{Float32, 1, Tuple{16}}
│ %18 = cuTile.Intrinsics.store_partition_view(%17, %13, nothing, nothing, %15, %20)::Nothing
└ return nothing
) => Nothing

After:

julia> ct.code_structured(vadd, Tuple{ct.TileArray{Float32, 1, ct.ArraySpec{1}(128, true, (0,), (32,))},
                                 ct.TileArray{Float32, 1, ct.ArraySpec{1}(128, true, (0,), (32,))},
                                 ct.TileArray{Float32, 1, ct.ArraySpec{1}(128, true, (0,), (32,))},
                                 ct.Constant{Int64, 16}})
1-element Vector{Pair{IRStructurizer.StructuredIRCode, DataType}}:
 StructuredIRCode(
│ %20 = cuTile.MakeTokenNode()::cuTile.TokenType()
│ %1  = cuTile.Intrinsics.get_tile_block_id(0)::cuTile.IntTile{Tuple{}, Int32}
│ %3  = cuTile.Intrinsics.make_tensor_view(_2)::cuTile.TensorView{Float32, 1}
│ %4  = cuTile.Intrinsics.make_partition_view(%3, (16,), cuTile.PaddingMode.Undetermined, cuTile.nothing)::cuTile.PartitionView{Float32, 1, Tuple{16}}
│ %6  = Core.tuple(%1)::Tuple{cuTile.IntTile{Tuple{}, Int32}}
│ %7  = cuTile.Intrinsics.load_partition_view(%4, cuTile.nothing, cuTile.nothing, %6, %20)::cuTile.FloatTile{Tuple{16}, Float32}
│ %8  = cuTile.Intrinsics.make_tensor_view(_3)::cuTile.TensorView{Float32, 1}
│ %9  = cuTile.Intrinsics.make_partition_view(%8, (16,), cuTile.PaddingMode.Undetermined, cuTile.nothing)::cuTile.PartitionView{Float32, 1, Tuple{16}}
│ %11 = Core.tuple(%1)::Tuple{cuTile.IntTile{Tuple{}, Int32}}
│ %12 = cuTile.Intrinsics.load_partition_view(%9, cuTile.nothing, cuTile.nothing, %11, %20)::cuTile.FloatTile{Tuple{16}, Float32}
│ %13 = cuTile.Intrinsics.addf(%7, %12)::cuTile.FloatTile{Tuple{16}, Float32}
│ %15 = Core.tuple(%1)::Tuple{cuTile.IntTile{Tuple{}, Int32}}
│ %16 = cuTile.Intrinsics.make_tensor_view(_4)::cuTile.TensorView{Float32, 1}
│ %17 = cuTile.Intrinsics.make_partition_view(%16, (16,), cuTile.PaddingMode.Undetermined, nothing)::cuTile.PartitionView{Float32, 1, Tuple{16}}
│ %18 = cuTile.Intrinsics.store_partition_view(%17, %13, nothing, nothing, %15, %20)::Nothing
└ return nothing
) => Nothing

Motivation: this does away with the existing scalar elimination pass and should make it easier to add rewrite rules (that now don't have to think about scalar vs tile).