[mlir][tosa] Introduce accumulator type for `reduce_sum` on bf16 (#158389)
TOSA requires that `reduce_sum` operations on bf16 accumulate into fp32.
This change updates the `linalg` legalization by introducing an explicit
accumulator type to ensure compliance with the specification.
---------
Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>