Context

While working on PYTHON-5965, I found a performance win that affects the existing
bson. I measured about a 15% improvement on a spawn perf host on the standard bson tests.

{write_unicode}} (string values) and decode_and_write_pair (string keys) allocate a temporary bytes per element using PyUnicode_AsUTF8String.

Definition of done

• Switch to PyUnicode_AsUTF8AndSize, which returns a borrowed pointer into the string's cached UTF-8 buffer with no temporary allocation.

Pitfalls

The pointer is borrowed, so keep the source string alive until after the buffer write (else use-after-free)