[DRAFT] Shared/aggregate load

[alefimov-amd]

Introduces AggregateLoad pass, which aggregates multiple small loads inside a loop in one wide load and moves it it outeerloop:

for i in range(0, 32):
  val = global_load: tensor<8x8>

transforms to:

for i in range(0,2):
  wide_load = global_load: tensor<8x128>
  smem = local_alloc(val)
  for i in range(0, 16):
    view = subview(smem)
    val = local_load(view): tensor<8x8>

In case all iterations could be aggregated, outer loop is not created.

For now application of this pass is limited to scale parameters of scaled_dot operation.