NNG Protocol Level Memory Enhancements

[shikokuchuo]

The mirai v2.0 architectural enhancement already improves memory usage, as we moved from:

host <-- `req/rep`--> dispatcher <-- `req/rep` --> daemon

to:

host <-- `req/rep`--> dispatcher <-- `poly` --> daemon

The poly protocol does not retain messages longer than necessary and is efficient in that regard.

This PR adds memory optimization for the remaining req/rep leg:

• The requestor on host still retains a copy of the message until the reply is received. This is even the case when we've turned off NNG's automatic protocol-level retry mechanism.
• The NNG source is patched to free the request message early as soon as the initial send is complete (as it won't be used again).
• Substantially lower memory usage on host if many mirai requests are outstanding (for example in a map operation).

The NNG author has previously agreed that this would be a sensible amendment to the protocol, so if everything tests out here, the patch will be proposed to upstream.

[shikokuchuo]

@wlandau This is an important one for your use in targets. Are you fine to do some testing based on this branch / PR?

If it's easier for you to grab R-universe builds, I'm ok to go ahead and merge this first.

[wlandau]

All the tests in https://github.com/wlandau/crew/tree/main/tests/testthat and https://github.com/wlandau/crew/tree/main/tests/local pass on Mac OS and RHEL9 using shikokuchuo/mirai@640468b and e38f178.

[shikokuchuo]

Proposed to upstream in nanomsg/nng#2107. Merging this PR.