multi: multiplexing improvements

Fixes #3436
Closes #3448

 Problem 1

After LOTS of scratching my head, I eventually realized that even when doing
10 uploads in parallel, sometimes the socket callback to the application that
tells it what to wait for on the socket, looked like it would reflect the
status of just the single transfer that just changed state.

Digging into the code revealed that this was indeed the truth. When multiple
transfers are using the same connection, the application did not correctly get
the *combined* flags for all transfers which then could make it switch to READ
(only) when in fact most transfers wanted to get told when the socket was
WRITEABLE.

 Problem 1b

A separate but related regression had also been introduced by me when I
cleared connection/transfer association better a while ago, as now the logic
couldn't find the connection and see if that was marked as used by more
transfers and then it would also prematurely remove the socket from the socket
hash table even in times other transfers were still using it!

 Fix 1

Make sure that each socket stored in the socket hash has a "combined" action
field of what to ask the application to wait for, that is potentially the ORed
action of multiple parallel transfers. And remove that socket hash entry only
if there are no transfers left using it.

 Problem 2

The socket hash entry stored an association to a single transfer using that
socket - and when curl_multi_socket_action() was called to tell libcurl about
activities on that specific socket only that transfer was "handled".

This was WRONG, as a single socket/connection can be used by numerous parallel
transfers and not necessarily a single one.

 Fix 2

We now store a list of handles in the socket hashtable entry and when libcurl
is told there's traffic for a particular socket, it now iterates over all
known transfers using that single socket.