shard_connection: Handle non-blocking I/O errors on Unix socket client connect

[LINKIWI]

Currently, the client socket is configured as non-blocking, but non-blocking I/O errors are not properly handled on connect(2). (Reference)

This seems to be problematic for Unix socket clients, where a restrictive server TCP listen backlog (or a restrictive system setting for net.core.somaxconn) relative to the number of concurrent memtier_benchmark clients (--clients) results in persistent benchmark execution failures with Resource temporarily unavailable.

A simple reproduction case:

# Artifically restrict the listen(2) backlog queue size.
$ echo -en 'port 0\nunixsocket /tmp/redis.sock\ntcp-backlog 1\n' | ./redis-server -
...

$ ./memtier_benchmark -S /tmp/redis.sock --clients 100
Writing results to stdout
[RUN #1] Preparing benchmark client...
connect failed, error = Resource temporarily unavailable
prepare: failed to connect, test aborted.
error: failed to prepare thread 0 for test.

This patch augments the connect(2) path to handle EINPROGRESS and EAGAIN/EWOULDBLOCK errors; the connect is simply retried indefinitely for the latter case.

With this patch:

$ ./memtier_benchmark -S /tmp/redis.sock --clients 100
Writing results to stdout
[RUN #1] Preparing benchmark client...
[RUN #1] Launching threads now...
^CUN #1 27%,   2 secs]  4 threads:     1078250 ops,  359004 (avg:  359421) ops/sec, 14.87MB/sec (avg: 14.89MB/sec),  1.11 (avg:  1.11) msec latency
...