TextIO can't read concatenated gzip files

$ cat <<END > header.csv
a,b,c
END
$ cat <<END > body.csv
1,2,3
4,5,6
7,8,9
END
$ gzip -c header.csv > file.gz
$ gzip -c body.csv >> file.gz

The file is well-formed:
$ gzip -dc file.gz
a,b,c
1,2,3
4,5,6
7,8,9

However, TextIO.Read.from("/path/to/file.gz") will read only "a,b,c" - reproducible even when the file is on local disk and with the DirectPipelineRunner.

The bug is in CompressedSource. It uses GzipCompressorInputStream, which by default reads only the first gzip stream in the file, but has an option to read all of them. Previously (in Dataflow SDK 1.4.0) we used GZIPInputStream which reads all streams.