Data-dependent sinks

Current sink API writes all data to a single destination, but there are many use cases where different pieces of data need to be routed to different destinations where the set of destinations is data-dependent (so can't be implemented with a Partition transform).

One internally discussed proposal was an API of the form:
```

PCollection PCollection.apply(
Write.using(DoFn where,
MapFn> how)

```

so an item T gets written to a destination (or multiple destinations) determined by "where"; and the writing strategy is determined by "how" that produces a WriteOperation (current API - global init/write/global finalize hooks) for any given destination.

This API also has other benefits:
* allows the SinkT to be computed dynamically (in "where"), rather than specified at pipeline construction time
* removes the necessity for a Sink class entirely
* is sequenceable w.r.t. downstream transforms (you can stick transforms onto the returned PCollection, while the current Write.to() returns a PDone)

Imported from Jira [BEAM-92](https://issues.apache.org/jira/browse/BEAM-92). Original Jira may contain additional context.
Reported by: jkff.

Start by reading the current sink API and Write.to() behavior described in the issue, then review the original Jira BEAM-92 for additional context. Define and implement the proposed data-dependent Write API so destinations can be computed per item and the returned PCollection can be used by downstream transforms; identify the relevant tests while tracing the existing sink implementation.

Written by the indexing model from the issue text.