fix: Optimize dynamic string building

[mwcampbell]

This reduces intermediate allocations, especially when dynamically building a label.

One could imagine taking advantage of this in a self-voicing adapter to write the label, value, or text range directly into the buffer that contains the utterance to be spoken, and that buffer could be reused across utterances.

On Windows, this allows us to construct the dynamic string as UTF-16 in one pass, rather than constructing it as UTF-8 and then re-encoding it to UTF-16 in another pass, adding a temporary allocation.

This change happens to add 512 bytes to the binary size of a panic_immediate_abort build of the Windows hello_world example on x86-64. It doesn't have any effect on the binary size of that same example built as a normal release build with panic = "abort".