Fixes #2292

Fixes #2291

This done via a new http_CollectHdrSep function, implying a VRT minor
bump.

Refs #2291

Newly loaded VCLs didn't get their go cold timer set even when they
were set to auto, causing them to never go cold automatically.

Fixes: #2270

Fixes #2289

Lck_CondWait() will stopped in the tracks.

This should fix some H2 tickets.

Insist we get the frames we expect.

mark up functions where only it can go.

Mark scheduled h2_req's.

Fixes #1875 (maybe)

than to delay a random-ish number of seconds

A probe may still be running when its backend is deleted. In that case
it needs to read the proxy_header field under the lock and give up if
the backend was already gone.

Not sure how to write a test case for this one.

Fixes #2278

backend if still in mid-transfer.  With sufficiently small -bodylen,
as in this test, that might or might not happen.

Explicitly close from the server side, to stabilize this test.

Fixes #2280

On systems where both 32 and 64-bit libraries can co-exist, simply
inheriting Varnish's prefix is not enough because VMODs might land
in the wrong directory. For instance, on rhel you will find both
/usr/lib and /usr/lib64.

This is usually not a problem with package managers, because they
usually take care of everything at configure time, for example RPMs
may use a %configure macro that defines all well-known directories.
It can however be overlooked when a VMOD is manually built from
source.

This reverts the patch for #2207, which as martin points out, is not
a good idea after all.

Original patch from @joshuaspence.

Closes #2262

"leave this much free in workspace".

Add debug.typesize() which can tell us the C-sizeof() various types.

This enables waiting for the logging thread to idle to make sure all
generated log is caught.