fix(arc): Per-CID locking. Sync Map to CID lock

[frrist]

No description provided.

[gammazero]

Having per-cid locking may be slightly more complex, and there could be some additional overhead with allocating locks and growing a map. For your consideration, here is a very simple alternative that hashes keys over a fixed set of locks:
https://github.com/kubernetes/utils/tree/master/keymutex

If you do prefer per-cid locking, allow me to suggest that you do not need a separate mutex for each cid. The presence of a map entry (a struct{}{}) is sufficient to determine if a resource is locked. If the cid is already locked (map entry present), Then you can use sync.Cond.Wait() to wait for a signal to check again to see if the resource is unlocked (map entry gone). Lock the resource by creating the map entry, unlock by removing it and signaling the sync.Cond.
Example here: https://github.com/gammazero/kmutex/blob/master/kmutex.go
That example uses interface{} as its key type, but there was no performance difference when benchmarked against ints and strings.

The things here are not meant to prescribe a specific implementation, but to be helpful examples if considering different implementations.

[Stebalien]

Closing based on #64 (comment).