Add gzip and zlib support for FixedLengthRecordReader (#8901)

* Switch from InputBuffer to BufferedInputStream for FixedLengthRecordReader

This commit switch from InputBuffer to BufferedInputStream for
FixedLengthRecordReader, for the purpose of allowing compressed file as
FixedLengthRecordReader.

Signed-off-by: Yong Tang <yong.tang.github@outlook.com>

* Remove the need to file the file size before hand for FixedLengthRecordReader

This commit updates the implementation of FixedLengthRecordReader so that
it will not try to find the file size before hand.

The purpose is to allow FixedLengthRecordReader to take a non-seekable file
as an input. In a non-seekable file (e.g., a compressed file), it may not
be possible to find the file size (e.g., in a uncompressed file length).

This commit is part of the effort to support gzip/zlib compression.

Signed-off-by: Yong Tang <yong.tang.github@outlook.com>

* Add gzip and zlib support for FixedLengthRecordReader
This fix adds gzip and zlib support for FixedLengthRecordReader,
as was discussed in 8856.
When FixedLengthRecordReader is used, it will check for
encoding flag and use ZlibInputStream as needed.
The usage of InputBuffer in FixedLengthRecordReader has
also been changed to BufferedInputStream to match ZlibInputStream.

This fix fixes 8856.

Signed-off-by: Yong Tang <yong.tang.github@outlook.com>

* Add additional test cases where num_records=0 and hop_bytes>record_bytes

This commit adds additional test cases where num_records=0 and hop_bytes>record_bytes

This commit also updated the API golden so that API compatibility test passes.

Signed-off-by: Yong Tang <yong.tang.github@outlook.com>