Go SDK: faster transforms/filter.Distinct with CombinePerKey

The current implementation:

1. add fixed value 1: P<T> --> P<<T, 1>>
2. group by key: P<<T, 1>> --> GBK<T, 1>
3. drop the value: P<distinct T>

The new proposed implementation:
1. ditto
2. combine by key: P<<T, 1>> --> P<<distinct T, 1>>
3. ditto

CombinePerKey performs a pre-GBK ParDo, which is useful to reduce the shuffle size.