Consider linalg.index an affine dim

[cferry-AMD]

llvm#118478 changed the definition of what a symbol is: it made it so that an operation that is Pure and whose operands are valid symbols, returns a valid symbol as result. This causes linalg.index, that matches this definition, to become a symbol -- contradicting its lowering to a loop counter by another pass convert-linalg-to-affine-loops.
Example (introduced as a test by this PR):

#set = affine_set<(d0) : (-d0 + 5 >= 0)>

func.func @linalg_index_affine_if(%in: tensor<10xf32>) -> tensor<10xf32> {
  %empty = tensor.empty() : tensor<10xf32>
  %out = linalg.generic {
      indexing_maps = [affine_map<(i) -> (i)>, affine_map<(i) -> (i)>], 
      iterator_types = ["parallel"]}
    ins(%in : tensor<10xf32>)
    outs(%empty : tensor<10xf32>) {
      ^bb0(%a: f32, %b: f32):
        %c42f = arith.constant 42.0 : f32
        %i = linalg.index 0 : index
        %ret = affine.if #set(%i) -> f32 {
          affine.yield %a : f32
        } else {
          affine.yield %c42f : f32
        }
        linalg.yield %ret : f32
    } -> tensor<10xf32>
  return %out : tensor<10xf32>
}

The upstream behavior canonicalizes the set (attribute of the affine.if) into:

#set = affine_set<()[s0] : (-s0 + 5 >= 0)>

where with the current PR, it stays as

#set = affine_set<(d0) : (-d0 + 5 >= 0)>

This PR allows to directly substitute linalg.index operands to affine ops with loop counters, without causing an error like operand cannot be used as a symbol due to how the map/set gets canonicalized with linalg.index.

Once downstream no longer needs it, this PR should be reverted, and let upstream take their stance between Affine and Linalg compatibility.

This PR also reverts the changes from upstream affecting the tests where linalg.index is used (all the tests touched by this PR except affine-if.mlir).