Annotate ModuleOp scheme/backend attribute

[ZenithalHourlyRate]

Follows #1420 (comment). The goal is making an "unmanaged" pipeline for openfhe so the loop support can work without all the mgmt nuances.

To do that I used the least managed strategy for B/FV and reuse it for BGV/CKKS when encountered. To get the multiplicativeDepth property for openfhe, the MulResultAnalysis becomes MulDepthAnalysis, which is different from level analysis as the former relies on mulop and the later relies on modreduce.

Parameter generation will just generate the fallback parameter when seeing the openfhe flag.

[ZenithalHourlyRate]

Though for the current loop support we will run into assertion...

bazel run //tools:heir-opt -- $PWD/tests/Examples/common/halevi_shoup_matmul.mlir --annotate-module="backend=openfhe" --mlir-to-ckks="ciphertext-degree=16 experimental-disable-loop-unroll=true"

The IR can survive all the way through linalg-to-tensor-ext and mgmt passes, and get

    %0 = secret.generic ins(%arg0 : !secret.secret<tensor<1x16xf32>>) attrs = {__argattrs = [{mgmt.mgmt = #mgmt.mgmt<level = 1>}], __resattrs = [{mgmt.mgmt = #mgmt.mgmt<level = 1>}]} {
    ^body(%input0: tensor<1x16xf32>):
      %1 = arith.mulf %input0, %cst {mgmt.mgmt = #mgmt.mgmt<level = 1>} : tensor<1x16xf32>
      %2 = arith.addf %1, %cst_1 {mgmt.mgmt = #mgmt.mgmt<level = 1>} : tensor<1x16xf32>
      %3:2 = affine.for %arg1 = 1 to 16 iter_args(%arg2 = %2, %arg3 = %input0) -> (tensor<1x16xf32>, tensor<1x16xf32>) {
        %4 = tensor_ext.rotate %arg3, %c1 {mgmt.mgmt = #mgmt.mgmt<level = 1>} : tensor<1x16xf32>, index
        %extracted_slice = tensor.extract_slice %cst_0[%arg1, 0] [1, 16] [1, 1] : tensor<16x16xf32> to tensor<1x16xf32>
        %5 = arith.mulf %4, %extracted_slice {mgmt.mgmt = #mgmt.mgmt<level = 1>} : tensor<1x16xf32>
        %6 = arith.addf %arg2, %5 {mgmt.mgmt = #mgmt.mgmt<level = 1>} : tensor<1x16xf32>
        affine.yield %6, %4 : tensor<1x16xf32>, tensor<1x16xf32>
      } {__argattrs = [{mgmt.mgmt = #mgmt.mgmt<level = 1>}, {mgmt.mgmt = #mgmt.mgmt<level = 1>}], mgmt.mgmt = #mgmt.mgmt<level = 1>}
      secret.yield %3#0 : tensor<1x16xf32>
    } -> !secret.secret<tensor<1x16xf32>>

Which is certainly something Openfhe is happy with. Then DistributeGeneric is unhappy somewhere

heir-opt: lib/Dialect/Secret/Transforms/DistributeGeneric.cpp:345
Assertion `isa<BlockArgument>(value) && "not sure what to do here, file a bug"' failed.

I think this should be addressed in another issue and should not affect this PR. The loop produced by this PR is valid.

[j2kun]

I think there may be some places where the top level getOperation() in the pass is not pointing to a module op (e.g., if you run a pass pipeline that anchors on func.func) but this can be fixed when we run into it.