The worker workspace gets reset in miss/pass/pipe, so a copied substring
could get overwritten after recv.

This should fix an assertion failure when the ov became invalid between
a match and backref.

  match (according to the documentation). Change implementation and
  test to follow the documentation
- don't keep pcre internal state in struct sess_ov
- clean up struct sess_ov if the xid changes and if the workspace
  changes
- stramline getting the struct sess_ov

null byte of the back reference copy to the workspace, so any data from
following workspace allocations could appear to be appended to the return
value of re.backref() / vmod_backref.

The previous code had some issues:

- race: priv_call->priv could be incomplete when being non-null, so
  using it outside re_mutex was not safe.

  fix: make sure that the re_t is complete before referencing it in
  priv_call->priv using a membar

- caching dynamic REs did not work because invalid comparison of
  the cached pattern, also freeing the cached REs was incorrect

  it could be a good idea to add a global cache for compiled
  patterns (dynamic or not), but as long as we don't have that,
  it's probably safest to just recompile dynamic REs always

always include config.h first -- solaris (and others?) need defines from it (GNU_SOURCE in particular)

compilation fails).

and pattern.

(because we don't have any, just varnishtest)

succeeds, and has not already been done in another thread

always return true for matches against the empty pattern