FTP(S) use two connections that can be set to different recv and
send functions independently, so by introducing recv+send pairs
in the same manner we already have sockets/connections we can
work with FTPS fine.

This commit fixes the FTPS regression introduced in change d64bd82b.

Howard Chu brought the bulk work of this patch that properly
moves out the sending and recving of data to the parts of the
code that are properly responsible for the various ways of doing
so.

Daniel Stenberg assisted with polishing a few bits and fixed some
minor flaws in the original patch.

Another upside of this patch is that we now abuse CURLcodes less
with the "magic" -1 return codes and instead use CURLE_AGAIN more
consistently.

  a broken TLS server. However it does not happen if SSL version is selected
  manually. The approach was originally taken from PSM. Kaspar Brand helped me
  to complete the patch. Original bug reports:
  https://bugzilla.redhat.com/525496
  https://bugzilla.redhat.com/527771

  closed NSPR descriptor. The issue was hard to find, reported several times
  before and always closed unresolved. More info at the RH bug:
  https://bugzilla.redhat.com/534176

  in non-blocking mode.

false positive on a leaked socket, so this introduces a way to tell the system
that the socket is indeed closed without explicitly closing it!

  libcurl called NSS to close the SSL "session" it also closed the actual
  socket.

added support for new SQLite cert database format: added a runtime check for version 3.12.0, and depending on the result add 'sql:' prefix to cert database directory so that newer SQLIte database format works.

added aditional check for the directory specified with SSL_DIR, and fall back to hardcoded directory if not a valid directory.

  libcurl-NSS. Originally reported at:
  https://bugzilla.redhat.com/show_bug.cgi?id=516056#c9

  verbose error messages. Originally reported at:
  https://bugzilla.redhat.com/show_bug.cgi?id=516056

  in NSS-powered libcurl. Now the client certificates can be selected
  automatically by a NSS built-in hook. Additionally pre-login to all PKCS11
  slots is no more performed. It used to cause problems with HW tokens.

- Fixed reference counting for NSS client certificates. Now the PEM reader
  module should be always properly unloaded on Curl_nss_cleanup(). If the unload
  fails though, libcurl will try to reuse the already loaded instance.

  issue with client certs that caused issues like segfaults.
  http://curl.haxx.se/mail/lib-2009-05/0316.html

  PK11_CreateGenericObject() function.

  at https://bugzilla.redhat.com/show_bug.cgi?id=453612#c12

  If an incorrect password is given while loading a private key, libcurl ends
  up in an infinite loop consuming memory. The bug is critical.

  out to be leaking cacerts. Kamil Dudka helped me complete the fix. The issue
  is found in Redhat's bug tracker:
  https://bugzilla.redhat.com/show_bug.cgi?id=453612

  There are still memory leaks present, but they seem to have other reasons.

  NSS is used. These ciphers were added in NSS 3.4 and require to be enabled
  explicitly.

  I have to jump through a few hoops now with the NSS library initialization
  since another part of an application may have already initialized NSS by the
  time Curl gets invoked. This patch is more careful to only shutdown the NSS
  library if Curl did the initialization.

  It also adds in a bit of code to set the default ciphers if the app that
  call NSS_Init* did not call NSS_SetDomesticPolicy() or set specific
  ciphers. One might argue that this lets other application developers get
  lazy and/or they aren't using the NSS API correctly, and you'd be right.
  But still, this will avoid terribly difficult-to-trace crashes and is
  generally helpful.

codes for all calls to malloc and strdup that were missing. I also changed
a few malloc(13) to use arrays on the stack and a few malloc(PATH_MAX) to
instead use aprintf() to lower memory use.
I also fixed a memory leak in Curl_nss_connect() when CURLOPT_ISSUERCERT is
in use.