S3Store#put uses unbounded memory when `Path.size` is None

[Daenyth]

S3Store invokes TransferManager#upload, which says

     * When uploading options from a stream, callers <b>must</b> supply the size of
     * options in the stream through the content length field in the
     * <code>ObjectMetadata</code> parameter.
     * If no content length is specified for the input
     * stream, then TransferManager will attempt to buffer all the stream
     * contents in memory and upload the options as a traditional, single part
     * upload. Because the entire stream contents must be buffered in memory,
     * this can be very expensive, and should be avoided whenever possible.

When you use too much, you get a java.lang.OutOfMemoryError: Java heap space

It is possible to do multipart uploads to use fixed memory; for reference, this is what Alpakka does for akka-streams:

• Alpakka S3Client#multipartUpload
• Using the AWS Java SDK for a Multipart Upload (Low-Level API)

  • Use the low-level API when you [..] do not know the size of the upload data in advance

[Daenyth]

I have a draft solution for this here; it fits an internal api we're using for the moment but once I get it tested I should hopefully be able to PR up the changes.

https://gist.github.com/Daenyth/3ca1858c35b4a2daf5ea04ed8bf6800e