Instead of having a single mmap(2)'ed file inside which we allocate
the necessary chunks, put each chunk in its own file and use two index
files (MGR/CHILD) to keep track of them.

Major benefits:

* You don't need to guess how much VSM space you need at startup.

* VSM chunk Identifier (and therefore VSC counter names) are unlimited length.

Minor benefits:

* VSM API has simpler and explainable semantics now.

* The bogo-you-didnt-guess-right-malloc-VSM allocations are gone.

* -t argument moves up/down to VSM level and works the same through-out.

* VSM/VUT utils now stay attached even if the MGT process restarts.

Some relatively minor adjustments are still to come before the VSM API is frozen
and the .map file finalized for the release.

Previously, we could run out of stack handling stack overflows, leaving
users with unspecific SIGSEGV crashes and no panic message.

By providing a single alternative stack exclusively for SIGSEGV handling
where sigaltstack() is available, we increase chances for our signal handler
to finish successfully.

In particular, this will make it easier to diagnose stack overflows by
comparing the failing address with the stack info from the panic output.

This could be further improved by giving advise to increase thread_pool_stack
if si_addr is near the stack boundaries.

c00057.vtc now triggers a stack overflow instead of raising a SIGSEGV.

Merges #2396

it was already used for more than SIGSEGV, so we should output the actual
signal description.

For HitMiss and HitPass, only show the remaining TTL. Seeing these
values should help debugging of timing-sensitive transactions. The
grace and keep periods are absolute, but the TTL is the remaining
time relative to the current request. Same as obj.{ttl,grace,keep}
variables in VCL.

The former is work in progress.

This is available after vcl_hash{}. Tests will follow.

LSAN is still disabled for now.

Travis sets CC _after_ before_install so use a different variable.

Reported by LSAN.

two '-' prefix to be sure to apply where we want it.

Spotted while playing with: PC-lint Plus (RC1)

Reported by LSAN.

Reported by LSAN.

Otherwise a field goes missing in SessOpen records.

When the do..while loop in HSH_Purge executes on a oh with many
popular variants, there is a potential problem with the "array" of oc
pointers, allocated in the thread workspace. If many of the oc's have
positive refcounts, they will fill up the array and

	EXP_Rearm(oc, now, ttl, grace, keep);
	(void)HSH_DerefObjCore(wrk, &oc, 0);

will be called several on the same oc's. At the same time, the counter
n_obj_purged will be updated with a too low number.

The test case demonstrates how we get a too low value for this counter,
but it is not able to force varnishd to use a siginificant amount of CPU.

This lets the request's vfp stack survive a VCL label change.

Fixes: #2367

Spotted by @fgsch

Refs #2362