Implement shared-memory communication

[mcopik]

Serverless functions are becoming more powerful, and AWS Lambda now supports using up to 6 vCPUs per function. We could benefit by running multiple "workers" as threads on a single function and letting them communicate locally. We cannot use processes within a single function as that would require shmem, which is not supported on AWS Lambda.

• Introduce a memory-based communication channel.
• Adjust the function launcher to support a hierarchy of functions - local and external ones.
• Implement in collectives and P2P communication the selection of external and local channels based on the recipient.

This might require making collectives to be natively multi-protocol first #6.