Add support for appending to files

[pawalt]

Tell us more about this new feature.

One common use case for S3 is storing log data. In a traditional disk-based system, you'd open a file in append mode, append data, and close it once you're done. You can pick up the append later on when you have more logs.

With mountpoint (and S3 in general), this is not possible. Because all files must be open with O_TRUNC, appending is impossible. This means that mountpoint doesn't play nicely with logging systems like Tensorboard which only support append-based logging.

Notably, this is not an ask because of efficiency - it's an ask for compatibility with existing libraries. An append implementation that pulled the whole object out of S3 and wrote the whole object back on flush() would be acceptable.

cc @thundergolfer