Remove the calls to `recvMalloc` in favour of using the (patched) `recv`
from the original nanomsg library, which shouldn't segfault anymore. The
reason for using `recv` are a few, but mostly the fact that `recv` can
allocated arbitrary-long payloads data (up to the 1MB limit) without an
hardcoded limit like `recvMalloc` was imposing. Furthermore, `recv` does
resource cleanup for us via `c_nn_freemsg`, whereas `recvMalloc` is not
thread/exception safe. Consider the implementation:

```
recvMalloc :: Receiver a => Socket a -> Int -> IO ByteString
recvMalloc (Socket t sid) numBytes = do
  ptr <- mallocBytes numBytes
  -- receive by blocking the thread
  len <- c_nn_recv sid ptr (#const NN_MSG) 0 -- (#const NN_DONTWAIT)
  str <- C.packCStringLen (castPtr ptr, fromIntegral len)
  free ptr
  return str
```

If any exception (synchronous or asynchronous) strikes _before_ the call
to `free`, we would be leaking C memory.

This makes it easier to change protocols, ports etc.

Also, will use stm containers

I'm currently able to do the following:
- start gargantext-server (as it hosts central exchange and
dispatcher, currently)
- start a websocket connection:
  websocat ws://localhost:8008/ws
- subscibe to a topic (in websocat):
  {"request": "subscribe", "topic": {"type": "update_tree", "node_id":
  15}}
- optionally subscibe to other node_ids or start other websocat's with
  different subscriptions (can be multiple)
- fire up
  cabal v2-run gargantext-central-exchange -- client
  This triggers a node_id: 15 and node_id: 16 notification to be sent

You can send your own notifications, e.g. with Python:
import json
import nanomsg as n
s = n.Socket(n.PUSH)
s.connect('tcp://localhost:5560')
s.send(json.dumps({'type': 'update_tree_first_level', 'node_id': 15}))