Many blobs aren't shared. Save 8 bytes per SharedBlob by using a normal
unordered_map instead of instrusive::set.
More importantly, perhaps, it avoids us having to tune the intrusive
unordered_set size manually. std::unordered_map does this automatically
for you, but the intrusive one does not. And it's unclear how to
statically size it given that it's a per-collection structure and we have
no idea how many objects we'll have, how many blobs per object, and how
many objects will be cloned.
Signed-off-by: Sage Weil <email@example.com>