Optimize allocations during metadata extraction

[REQUIRED] Use case description

My music app relies on ExoPlayer's exposed metadata parsing (MetadataRetriever) to do mass scans of music libraries.

However, there seems to be some poor memory use during certain parts of metadata extraction, with entire ID3 tags and vorbis comments being directly read into memory.

When my metadata system encounters large cover art data and generally large tags, these repeated allocations and copies often slow down the process due to the constant GCs required.

Proposed solution

Refactor the container extractors to minimize reading out data into memory until the actual tag data is needed. For example, Id3Decoder would read from an input stream up until the identified bound of the tag, rather than reading from an in-memory ByteArray. Using some kind of composable reader pattern could be helpful if data transformation is required.

Alternatives considered

Using another tagging library that allows me to stream in this data. I don't like the duplication of container parsing, but can tolerate it for efficiency.