This will make counting leaked objects, and maintaining internal data structures of the leak detector, much faster.

The IdentityHashSet is based on open-addressing 2-choice hashing with 8-element buckets.
Conceptually, this is like bucketed cuckoo-hashing without the cuckoo-ing.
The 8-element buckets, and the addition of an 8-element stash for unresolvable conflicts, allow the table to have high utilisation.
The table array is a power-of-two size, so in practice utilisation will be between 45% to 90+%.
This means it'll take up less memory than the IdentityHashMap-as-a-Set we had before, in part due to not storing values, and in part to using a higher-utilisation hashing algorithm.
Theoretically, the 2-choice hashing scheme have higher access costs, due to computing two hashes and having at least two cache-misses, one from each sub-table, but this doesn't seem to be all that noticeable in the grand scheme of things.
The access performance is dramatically better than the 2-level chaining Bagwell trie we used before, by simply scaling to larger sizes.
The Bagwell trie, as implemented, didn't scale to more than 1000 elements, and relied entirely on chaining for capacities beyond that.

The `toString` was recursive, which is a problem when there are a large number of nodes.

This is useful when Lilliput will start requiring object header inflation for storing identity hash codes.
Identity hash codes are already somewhat expensive to compute on the first call.
Identity hash codes might also not have full 32-bit entropy, which can produce a poor distribution.
This method makes it possible to use alternative pre-computed hash codes, such as a process-wide counter.

…references in the precise leak detector

The added int field hides in alignment padding on most JVM configurations, so the objects won't actually take up more space from this added field.