Scatter Optimizations with Window Dims specified

[giordano]

julia> using Reactant
AssertionError("Could not find registered platform with name: \"cuda\". Available platform names are: ")

julia> x = Reactant.to_rarray([3])
1-element ConcreteRArray{Int64, 1}:
 3

julia> T = Any[nothing]
1-element Vector{Any}:
 nothing

julia> function ip(m, T)
           @allowscalar m[1] = 2
           T[1] = m
           return m
       end
ip (generic function with 1 method)

julia> @code_hlo ip(x, T)
module {
  func.func @main(%arg0: tensor<1xi64>) -> tensor<1xi64> {
    %c = stablehlo.constant dense<2> : tensor<1xi64>
    %c_0 = stablehlo.constant dense<0> : tensor<1x1xi64>
    %0 = "stablehlo.scatter"(%arg0, %c_0, %c) <{scatter_dimension_numbers = #stablehlo.scatter<inserted_window_dims = [0], scatter_dims_to_operand_dims = [0], index_vector_dim = 1>}> ({
    ^bb0(%arg1: tensor<i64>, %arg2: tensor<i64>):
      stablehlo.return %arg2 : tensor<i64>
    }) : (tensor<1xi64>, tensor<1x1xi64>, tensor<1xi64>) -> tensor<1xi64>
    return %0 : tensor<1xi64>
  }
}

[wsmoses]

There's a few ways this can be done

1. Scatter -> dynamicupdateslice
2. scatter of whole size -> replace (should be subsumed by 1)

[avik-pal]

1. Scatter -> dynamicupdateslice

we have this. we need to make sure it supports inserted_window_dims = [0]. (we could also potentially fix this on reactant end to not emit inserted_window_dims in the first place)

[wsmoses]

eh ideally we can optimize this (better long term anyways)