Select transform (Schema-based) result order is not predictable

pipeline.apply(Select.fieldNames("x", "y"))

pipeline.apply(Select.fieldNames("a", "b"))

The returned output order is not predictable. In the above two examples, field `x` may return first, while field `a` (also queries in the first place) may return in the second place.  

Shall we add `withOrderByFieldInsertionOrder` to fieldAccessDescriptor in Select PTransform, so that the return order is predictable?

Imported from Jira [BEAM-9016](https://issues.apache.org/jira/browse/BEAM-9016). Original Jira may contain additional context.
Reported by: yangzhang.