The http_PutProtocol() and http_PutResponse() would in the case of
workspace overflow leave the headers as NULL and log a
SLT_LostHeader. This would make Varnish assert correctly later when
writing to the wire, as these are mandated by HTTP.

This commit changes them to set the fields to static strings instead
("HTTP/1.1" and "Lost Response") when failing to write them to the
workspace. This leaves enough information to complete the protocol in the
case of overflow.

The patch also increases the synthetic object's workspace from static
1024 to param->http_resp_size. This leaves more (and configurable)
room for manipulating the headers of the synthetic object in
vcl_error.

This whole thing has been a collaboration between Martin and myself. I'll
leave it a mystery who wrote what line of code, which part of the comment
and contributed what to the test-case.

In all fairness, it's not a prefect solution, but a far step closer to one.
So it sort of, kinda, more or less, for now, until we get a better
solution:

Fixes: #1031

Conflicts:

	bin/varnishd/cache_http.c