use custom allocator/memory pool for `Token` allocations

The new/delete calls for the Token allocations make up more than 30% of the Ir count when profiling with callgrind.

It might make sense to implement a custom allocator or memory pool to reduce that.