Reduce memory usage during message encoding / serialization

[shikokuchuo]

Closes #219.

This PR allocates a nng_msg upfront and uses nng_msg_append() directly in all the internal encoding and serialization functions to avoid allocation of an intermediate object and memcpy - reducing peak memory usage.

Previously, we used a nano_buf struct (similar to base R's membuf). This held the serialized object before being appended in its entirety to an nng_msg. This would perform a memcpy() on the object - a real one, not one that would be optimized away by the compiler.

Note: we still retain the use of nano_buf in places such as stream sends which are not nng_msg based.

FYI @lionel-

[shikokuchuo]

Valgrind run comes in clean.

[shikokuchuo]

Rhub run also clean: https://github.com/r-lib/nanonext/actions/runs/19907841890