writeDynamic on EMR/Spark is not finishing writes to S3, leaves out temp files

Filing a bug as suggested on [StackOverflow](https://stackoverflow.com/questions/67163902/writedynamic-in-apache-beam-on-emr-spark-is-not-finishing-writes-to-s3-leaves-o?noredirect=1#comment118755306_67163902).

I have a bounded PCollection and would like to persist the output to S3 bucket with dynamic file naming scheme. Unfortunately, when running on EMR & Spark Runner (tried emr-6.2.0/Spark 3.0.1/Beam 2.28.0 and emr-5.30.1/Spark 2.4.5/Beam 2.14.0), after all steps in the pipeline finish and cluster terminates, output on S3 contains only some of expected contents in main output directory, most of it though is placed in .temp-beam folder and never moved to the main output dir. Most of it = 90% of expected lines are not persisted in correctly named files and spot checks indicate the expected lines are in files inside .temp-beam folder.
Here's a relevant pipeline declaration part:
```

PCollection input; // is a bounded PCollection

FileIO.Write write
= FileIO.writeDynamic()
.by(SomeObject::key)
.withDestinationCoder(StringUtf8Coder.of())

.withCompression(Compression.GZIP)
.withNaming((SerializableFunction)
key
-> (FileIO.Write.FileNaming) (window, pane, numShards, shardIndex, compression)
-> String.format("some_object_%s_%d.csv.gz",
key, shardIndex))
.via(Contextful.fn(SomeObject::toCsvLine), Contextful.fn(x -> TextIO.sink().withHeader(SomeObject.HEADER)))

.to("s3://some-bucket/some-output-path");
input.apply("write-a-pcollection", write);

```

With this code I get an S3 bucket that looks like:
* some_object_key1_0.csv.gz
* some_object_key1_1.csv.gz
* some_object_key2_0.csv.gz
* some_object_key3_0.csv.gz
* .temp-beam-\ with 90% of the expected content persisted inside objects named with random uuids

However, when I add `.withIgnoreWindowing()` to the writeDynamic configuration, output seems to be fully correct and no .temp-beam directory is left over. This method is deprecated though and no replacement has been ever provided in the javadocs (at least I couldn't find any).

Input PCollection does not contain any windowing and it does not solve the issue, when I force global windowing right before the writeDynamic transform.

This might be a bug in Beam or a consideration to leave out `.withIgnoreWindowing()` in the future versions. Kindly please investigate the issue.

Imported from Jira [BEAM-12202](https://issues.apache.org/jira/browse/BEAM-12202). Original Jira may contain additional context.
Reported by: pawelw.