default options:

For co-existance with libc malloc and potentially other allocators, we
should use mmap() regions in favor of sbrk(). Also, libumem's perthread
cache only makes sense when umem is actually used for all allocations
and, by default, requires an additional massive 1MB per thread, which
is totally out of scale with how we manage memory in varnish.

if libumem is found to already be loaded, we cannot set its options,
so warn about that fact.

As we only use function pointers to libumem, lazy loading of symbols
should be the better option.