This reverts commit bf2b7a78.

There are actually two off-by-one errors, one in WS_Assert_Allocated(),
and another in the use of -1 to have WS_Assert_Allocated() compute
the size of the string, including the null terminator.

We don't need the subject with the null terminator in the ov_t
structure (PRIV_TASK data), because we just need it to get backrefs
from the indices in the ov vector returned from pcre.

Also, there's a system in the VTC tests that may not be obvious.
The b*.vtc tests are about regex matching, c*.vtc for backrefs.

and add a test for it

Fixes #3

Closes #2

The .failed() and .error() methods have been removed.

The .match_dyn method has been removed, in its place we have the
functions match_dyn and backref_dyn.

Closes #1

for example if there was an error initializing the object.

required since 81bc4cef2a5e651a9e34ae5400672a88663db7ef

ensure that it builds with make -j

and it does since 49712a5fae5919e6e333d7164b98ef986b149c62

the addresses used as values point to workspace or static memory,
which are free'd by other means

expect successful response status, but Content-Length may or may not
be set in the response

number "in the middle" does not capture any string, e.g. when
"(a|(z))(bc)" matches "abc", with no capture for backref 2

use them because they lint the output of varnishd -C, we won't bother.

This is a complete re-implementation, not just for Varnish 4, but also
because, for the state maintained to access the ov structure in backref,
it does not use a session table. Instead, regex objects have a pthread key
for thread-local access to the ov in the ctx workspace.