Instead of buffering all TX IQ on a ring buffer, they can be put in a message buffer directly, reducing the number of memcpys in the TX chain. 1. pre-allocate message in `trx_write` function ``` zmq_msg_t msg; zmq_msg_init_size(&msg, size); ``` 2. Convert IQ from `c16_t` in the txdata buffer to `cf_t` in the message 3. push message onto a thread-safe queue 4. in `poll_thread`, use the queue instead of the buffer whenever the opposite side requests IQ This should allow higher throughput on slower CPUs.