fix(expiration): don't add nottl items

[KyleSanderson]

There's a ton of performance work that can happen here. The expiration queue dominates my profiles on real-world workloads, but that's because it's not doing a swap and is instead always rewriting the head pointer.

I can try to take this on, if you'll take it upstream.

[swithek]

Thank your for the PR.

Omitting NoTTL items from the expiration queue is a good idea and this PR provides a decent starting point, but in the current state, it might produce some side effects. For example:

• What happens if an item is first set with a non-zero TTL and then there's an update of the same item that changes its TTL to NoTTL?
• There are other methods that depend on the expiration queue, for instance, the Len() method. Will it be able to handle NoTTL items not being in the queue?