Add encode_in_place() to encode messages in-place.

[de-vri-es]

This is the encoding counterpart of decode_in_place. It is slower than encode_buf, but on embedded platforms it may be more important to save memory than to be as fast as possible.

Encoding in-place is slightly more complicated than decoding, because the message will grow while encoding it, and we must prevent overwriting data that hasn't been encoded yet.

To do that, the encoding works backwards: encode the message back to front. And do to that, it firsts computes where the encoded message will end so it can start writing there.

An alternative approaches would be to move the entire message to the back of the buffer first, and then encode from the front as usual. However, I wanted to avoid copying the whole message. I'm assuming that computing the output length is faster than writing to RAM (especially on embedded CPUs).

I added fuzzing and some test cases to ensure the implementation is correct.

Added a benchmark to compare it with the existing encoders: on my laptop it's still significantly faster than encode_iter() and indeed slower than encode_buf():