crypto/tls: replace net.Pipe in tests with real TCP connections
crypto/tls is meant to work over network connections with buffering, not
synchronous connections, as explained in #24198. Tests based on net.Pipe
are unrealistic as reads and writes are matched one to one. Such tests
worked just thanks to the implementation details of the tls.Conn
internal buffering, and would break if for example the flush of the
first flight of the server was not entirely assimilated by the client
rawInput buffer before the client attempted to reply to the ServerHello.
Note that this might run into the Darwin network issues at #25696.
Fixed a few test races that were either hidden or synchronized by the
use of the in-memory net.Pipe.
Also, this gets us slightly more realistic benchmarks, reflecting some
syscall cost of Read and Write operations.