Fix tls_connection shutdown cleanup

[proxyles]

The SSL library maintains an internal table of CA certificates
(ssl_otp_cacertificate_db). This is supposed to be cleaned up when the
last connection using a certificate closes, however there's two problems
in R16B02 (and in the current master branch on github):

• When CA certificates are provided as binary blobs, rather than by
  filename (ie, #ssl_options.cacerts is set, but #ssl_options.cacertfile
  is undefined) the cleanup never occurs due to an incorrect pattern
  match in tls_connection:handle_trusted_certs_db/1. This causes the
  table to grow unchecked because each connection adds a new entry.
• When the process exits abnormally, tls_connection:terminate/1 is never
  called because the trap_exit process flag is not set and so similarly
  the table (and everything else cleaned in terminate/1, for that
  matter) is not cleaned up. This doesn't affect "normal" termination
  caused by the connection closing because terminate/1 is called
  explicitly from handle_sync_event/4, rather that relying on gen_fsm's
  automatic calling.

[OTP-Maintainer]

Patch has passed first testings and has been assigned to be reviewed

[IngelaAndin]

The trap_exit problem is already solved, the other part of the patch we want. Can you please rebase that part onto master? Also if you could provide a test case that would be great. I think this should have been communicated already, I do not now how come it is not in the pull request log. We will have to look into that.

[proxyles]

Merged