Remove SetOp operation

[pthomadakis]

The operation ta.set_op creates several difficulties in analysis and rewriting/conversion passes and needs to be removed.
Specifically, this operation mutates an SSA in place, thus, making it difficult to track chains of operations that affect the same operand or
replacing its uses.

Many passes try to avoid this problem by replacing the use of set_op's 1st operand with operand 0 in every operation after itself, but this will not always work.

An example where it wouldn't work is the following:

#map = affine_map<(d0, d1, d2) -> (d0, d1)>
#map1 = affine_map<(d0, d1, d2) -> (d1, d2)>
#map2 = affine_map<(d0, d1, d2) -> (d0, d2)>
#map3 = affine_map<(d0, d1) -> (d0, d1)>
module {
  func.func @run_comet_with_jit(%arg0: !ta.sparse_tensor<f64, i64, ?x?, d, unk, cu, unk>) {
    %0 = "ta.index_label"() : () -> !ta.index
    %1 = "ta.index_label"() : () -> !ta.index
    %2 = "ta.index_label"() : () -> !ta.index
    %c7 = arith.constant 7 : index
    %3 = "ta.spTensor_decl"() <{format = "CSR", temporal_tensor = false}> : () -> !ta.sparse_tensor<f64, i64, ?x?, d, unk, cu, unk>
    %4 = "ta.spTensor_decl"() <{format = "CSR", temporal_tensor = false}> : () -> !ta.sparse_tensor<f64, i64, ?x?, d, unk, cu, unk>
    %5 = "ta.mul"(%arg0, %arg0, %0, %1, %1, %2, %0, %2) <{MaskType = "None", formats = ["CSR", "CSR", "CSR"], indexing_maps = [#map, #map1, #map2], operandSegmentSizes = array<i32: 1, 1, 6, 0>, semiring = "plusxy_times"}> {__alpha__ = 1.000000e+00 : f64, __beta__ = 0.000000e+00 : f64} : (!ta.sparse_tensor<f64, i64, ?x?, d, unk, cu, unk>, !ta.sparse_tensor<f64, i64, ?x?, d, unk, cu, unk>, !ta.index, !ta.index, !ta.index, !ta.index, !ta.index, !ta.index) -> !ta.sparse_tensor<f64, i64, ?x?, d, unk, cu, unk>
    "ta.set_op"(%5, %3) {__beta__ = 0.000000e+00 : f64} : (!ta.sparse_tensor<f64, i64, ?x?, d, unk, cu, unk>, !ta.sparse_tensor<f64, i64, ?x?, d, unk, cu, unk>) -> ()
    %6 = "ta.elews_mul"(%3, %arg0, %0, %1, %0, %1, %0, %1) <{formats = ["CSR", "CSR", "CSR"], indexing_maps = [#map3, #map3, #map3], semiring = "noop_times"}> {__alpha__ = 1.000000e+00 : f64, __beta__ = 0.000000e+00 : f64} : (!ta.sparse_tensor<f64, i64, ?x?, d, unk, cu, unk>, !ta.sparse_tensor<f64, i64, ?x?, d, unk, cu, unk>, !ta.index, !ta.index, !ta.index, !ta.index, !ta.index, !ta.index) -> !ta.sparse_tensor<f64, i64, ?x?, d, unk, cu, unk>
    "ta.set_op"(%6, %4) {__beta__ = 0.000000e+00 : f64} : (!ta.sparse_tensor<f64, i64, ?x?, d, unk, cu, unk>, !ta.sparse_tensor<f64, i64, ?x?, d, unk, cu, unk>) -> ()
    %7 = "ta.reduce"(%4) : (!ta.sparse_tensor<f64, i64, ?x?, d, unk, cu, unk>) -> f64
    "ta.print"(%7) : (f64) -> ()
    return 
  }

  func.func @main() {
    %3 = "ta.spTensor_decl"() <{format = "CSR", temporal_tensor = false}> : () -> !ta.sparse_tensor<f64, i64, ?x?, d, unk, cu, unk>
    %c0 = arith.constant 0 : index
    %4 = "ta.dim"(%3, %c0) : (!ta.sparse_tensor<f64, i64, ?x?, d, unk, cu, unk>, index) -> index
    %c1 = arith.constant 1 : index
    %5 = "ta.dim"(%3, %c1) : (!ta.sparse_tensor<f64, i64, ?x?, d, unk, cu, unk>, index) -> index
    "ta.fill_from_file"(%3) <{filename = "SPARSE_FILE_NAME0", readMode = 2 : i32}> : (!ta.sparse_tensor<f64, i64, ?x?, d, unk, cu, unk>) -> ()
    call @run_comet_with_jit(%3) : (!ta.sparse_tensor<f64, i64, ?x?, d, unk, cu, unk>) -> ()
    return 
  }

  func.func private @quick_sort(memref<*xindex>, index)
}

Removing this operation would also help us raise the abstraction from memref to tensor in many passes/conversions.
However, several passes rely on its existence.