For the case of gzip included in plain response, the esi_level == 1 vdp
order was:

	pesi gunzip pesi_buf V2P(to_parent)

Yet we had assertions in place that pesi_buf always immediately follows
pesi.

The reason was that, for esi_level > 0, we would not push pesi_buf from
pesi init but rather from the transport, which was plain wrong: We
should delay any additional vdps in order to buffer the least amount of
data.

Working on this, I also noted that for the generic buffering case, our
assertion that pesi_buf is first, might be too strict. Now, any VDPs
before the buffer are being closed at esi_level > 1.

fixes this panic:

Assert error in vped_close_vdp(), vdp_pesi.c line 1182:
  Condition(vdpe->vdp == vdp) not true.

	...

Backtrace:
  0x43cf3b: /usr/sbin/varnishd() [0x43cf3b]
  0x4a01c2: /usr/sbin/varnishd(VAS_Fail+0x42) [0x4a01c2]
  0x7f3719306b63:
./vmod_cache/_vmod_pesi.4d9e0603bac2a2e2b2627f7fe90ff1d55d4759545517c869a5571f16636e230e(+0x8b63)
[0x7f3719306b63]
  0x7f371930a377:
./vmod_cache/_vmod_pesi.4d9e0603bac2a2e2b2627f7fe90ff1d55d4759545517c869a5571f16636e230e(+0xc377)
[0x7f371930a377]
  0x441eab: /usr/sbin/varnishd(CNT_Request+0x11ab) [0x441eab]
  0x7f3719308043:
./vmod_cache/_vmod_pesi.4d9e0603bac2a2e2b2627f7fe90ff1d55d4759545517c869a5571f16636e230e(+0xa043)
[0x7f3719308043]
  0x45c833: /usr/sbin/varnishd() [0x45c833]
  0x45ccf0: /usr/sbin/varnishd() [0x45ccf0]
  0x7f37e6d18e65: /lib64/libpthread.so.0(+0x7e65) [0x7f37e6d18e65]
  0x7f37e6a4188d: /lib64/libc.so.6(clone+0x6d) [0x7f37e6a4188d]
thread = (cache-worker)
pthread.attr = {
  guard = 4096,
  stack_bottom = 0x7f372c482000,
  stack_top = 0x7f372c502000,
  stack_size = 524288,
}
thr.req = 0x7f3601106020 {
  vxid = 25559391, transport = PESI_INCLUDE
  step = R_STP_TRANSMIT,
  req_body = R_BODY_NONE,
  restarts = 0, esi_level = 1,

	...

We pass on reqs from esi subrequests to the top request for delivery.
Doing so we need to give them the top requests's worker such that VDPs
requiring it be happy.

Package includes the most recent bugfix, and is compatible with
Varnish 6.2.1.

When VDP_DeliverObj() was not called, for example for a head request or a
return code which implies no response body, bytes_tree->npending == 0
was not true.

To avoid additional complications, we code the fact that the root node,
if used, is pending into the npending_private field which meant for this
purpose, but otherwise only accounts for nodes below it. Yet, this is
not implied anywhere, so this use case should be perfectly fine.

Also add a test for HEAD requests on an actual ESI object.

Note on possible alternatives: I do not think a solution at VDP init
time is possible because, after the vmod gets pushed from VCL, the
response status could still be changed with effects on whether or not a
response body is to be sent (e.g. changed from 200 to 204 after the VDP
is pushed). So our only chance is to handle the case when the VDP gets
called next after _init, which is just _fini.

Strict compatibility with Varnish 6.2.0.

To reduce risks concerning portability and compatibility, and to
avoid accidentally depending on non-standard features.

- Use WS_Reserve() instead of WS_ReserveAll().
- Add HSH_Cancel() to the "foreign" code.

when we prune the tree, we may the current node containing the link
to the next unpending node may get freed, so we need to iterate using
the _SAFE variant which saves it.

the code is correct in varnish-cache, the VDP_bytes(req, VDP_NULL, ...) calls do not return.

This reverts commit ff7d20a2.

because the ESI request needs to stay alive until all nodes below have
been delivered, now that we got tree pruning, we can take them from the
ESI requests workspace.

See long comment in src/node.c for more detail and evidence collection
for the tuneables.

once a subtree is completely delivered, we do still need the T_NEXUS at
the head of the (sub)tree for integrity (linkage to its siblings), but
anything below can be completely fini'd and freed.

this is a bad leftover from when we still used ALLOC_JOB/FREE_OBJ for
nodes and anyway we must not just free a node which is still linked in
an otherwise alive tree

During code restructuring earlier in
77e9559d I overlooked that VTIM_real()
may not be declared where debug output is used

solve by moving the accumulation to the destructor

it is in the way of my torture testing and does not smell like something
we have caused..