Adding _tie_weights() to prediction heads to support low_cpu_mem_usage=True (#29024)
* Adding _tie_weights() to prediction heads to support low_cpu_mem_usage=True
* Testing for the non-safe-tensors case, since the default is safe-tensors already
* Running fixup/fix-copies
* Adding accelerate annotations to tests