The previous implementation used only one BUDDY_REQS, so whenever one
segment allocation was fulfilled, other requests with lower priority
could "get through" and ultimately lead to bfa_alloc() failing to
complete.

By using two BUDDY_REQS, we now make sure to "keep out place in the
priority queue".

We also limit cramming not only by the available bitfield segment slots,
but also by a maximum of 4 (1/16th of the requested size) and yield when
a lower cram does not succeed to motivate LRU more to make room.

This has undergone a _lot_ of testing and has gone through many
iterations, which all have been squashed into this commit.

since https://github.com/varnishcache/varnish-cache/pull/3855/files
was merged a long time ago and we never supported 7.2 in a public
release.

based on insights with minimum memory testing, we need even more fine
grained priorities.

The main insight is that new objects get lower priorities than
rewrite-related allocations, but once they exist, their subordinate
requests for segments and metadata get higher priorities.

We now also re-assign the priority of waiting allocations.

DLECHK gets a top priority because it uses small allocations and is
vital for rewrite progress.

When we hit an FCO for which we could not lock the varnish object (oc),
we inserted that FCO at the end of the LRU list, and restarted at the
first. But if that FCO was the only one, we would find ourselves in an
infinite loop.

We now put a marker segment on the LRU for the time we need to let go of
the lru lock and continue at that marker.

The exp thread is also critical. It must not starve, otherwise lru
starves also because of hitting OC_F_DYING objects.

Unfortunately, this brings us above 4K, so we fill up with io structs

This way we have a better chance to call the actual return at a more
suitable time, for example outside a lock.

Ref #53

Fixes #52

Closes #50