spotted while bringing syncparty' bufio.[ch] in sync; if the requested len could be bigger than buf->len plus the buf_grow() increment, so we have to iterate. This is currently impossible since we're usign buf_write() only to add one byte.

Add a proper upper limit (4096 seems enough for now) and safely NUL-terminate the rbuf. The idea is to later drop the clt->buf static buffer and allocate it on demand only when we need to do (http) chunking.