Ref https://github.com/varnishcache/varnish-cache/pull/3287

- minimum ttl of 10s for HFM objects
- simplify the X-TTL code
- show is_esi() example

this can save calls into the brotli code.

Note: I thought we might be able to avoid the additional ring buffer
copy in the brotli encode by using a buffer of size 1 << lgblock, but the
brotli code calls RingBufferWrite() even if the input buffer holds all
of the data to be compressed.

The corresponding varnish API function arguments are signed, make our
struct match them.

... such that the unix jail is in effect: Then the accept socket is
created as root and vcache:varnish cannot open it for writing

the delay is not an optimal solution, a barrier would be much more
appropriate, but I have not found the right place

The VFP trunk now requires Varnish master since the commit that
changed OF_CHGGZIP to OF_CHGCE.

See the discussion in Varnish issue #2910. A 304 IMS beresp could
(weirdly) set Content-Encoding to something other than the result
of previous VFP filtering, in this case compressing the response
with gzip. Nevertheless, Varnish does the right thing when the
OF_CHGGZIP flag is set.

Set the OF_CHGGZIP flag to indicate that:
- The ETag in a client response must be weakened if the response
  body differs from the fetched response.
- Any Content-Encoding header in the fetched 304 response should
  be removed.

... otherwise vcl_init fails with varnish-cache master and also
we would leak on vcl destroy.

to libbrotlicommon because it will not be found implicitly

See also https://github.com/google/brotli/pull/740