Push more work down to sender threads

[jcarreira]

Right now there is a lot of work that is done from the time an application calls put on the object store to the moment the request arrives at the sender working thread.

Mainly:

• Faltbuffer creation
• Message creation on the Flatbuffer
• Create future
• Enqueue builder into send_queue

If this work was instead pushed completely down to the sending thread, more work would offloaded from the application to Cirrus. This could enable more parallelism between both.