BigQuery Partitioned table creation/write fails when destination has partition decorator

Following is the code that writes to BigQuery: 
```

BigQueryIO.writeTableRows()
.to(destination)
.withCreateDisposition(CREATE_IF_NEEDED)
.withWriteDisposition(WRITE_APPEND)

.withSchema(tableSchema)
.expand(tableRows);
```

Here's the destination's implementation: 
```

public TableDestination apply(ValueInSingleWindow input) {
String partition = timestampExtractor.apply(input.getValue())

.toString(DateTimeFormat.forPattern("yyyyMMdd").withZoneUTC());
TableReference tableReference = new
TableReference();
tableReference.setDatasetId(dataset);
tableReference.setProjectId(projectId);

tableReference.setTableId(String.format("%s_%s", table, partition));
log.debug("Will write to BigQuery
table: %s", tableReference);
return new TableDestination(tableReference, null);
}
```

When the dataflow tries to write to this table, I see the following message:
```

"errors" : [ {
"domain" : "global",
"message" : "Cannot read partition information from a table
that is not partitioned: :.$19730522",
"reason" : "invalid"
} ]
```

So, it looks like it's not creating tables with partition in the first place? 

Apache beam version : 2.2.0

Imported from Jira [BEAM-3501](https://issues.apache.org/jira/browse/BEAM-3501). Original Jira may contain additional context.
Reported by: darshanmehta2.

Start at BigQueryIO.writeTableRows and follow CREATE_IF_NEEDED handling for TableDestination values containing partition decorators. Reproduce the supplied partitioned-table destination and verify that creation and subsequent append succeed without the “not partitioned” error.

Written by the indexing model from the issue text.