* long term feature: templated generic vm
* long term feature: shared_ptr less locator
* long term feature: shared_ptr less resource cache
* custom allocators and EnTT allocator-aware in general (long term feature, I don't actually need it at the moment) - see #22
* debugging tools (#60): the issue online already contains interesting tips on this, look at it
* runner proposal: https://en.wikipedia.org/wiki/Fork%E2%80%93join_model https://slide-rs.github.io/specs/03_dispatcher.html
* work stealing job system (see #100)
  - mt scheduler based on const awareness for types
* meta: sort of meta view based on meta stuff to iterate entities, void * and meta info objects
* allow for built-in parallel each if possible
* allow to replace std:: with custom implementations
* remove runtime views, welcome reflection and what about snapshot?
* types defined at runtime that refer to the same compile-time type (but to different pools) are possible, the library is almost there
* add opaque input iterators to views and groups that return tuples <entity, T &...> (proxy), multi-pass guaranteed
* add fast lane for raw iterations, extend mt doc to describe allowed add/remove with pre-allocations on fast lanes
* registry.each<T...>(first, last) by iterators, entities/components guaranteed
* multi component registry::remove and some others?
* built-in support for dual (or N-) buffering
* allow for custom stomp functions
* deprecate/replace snapshot
* remove dependency
* remove prototype

TODO
* custom (decoupled) pools ==> double buffering, shared components, multi-model
* make meta work across boundaries
  - inline variables are fine here, only the head represents a problem
  - we should always resolve by looking into the list of types when working across boundaries, no direct resolve
* nested groups: AB/ABC/ABCD/... (hints: sort, check functions)
* use unordered_map for named pools and context variables:
  * use direct access (pool-like) also for context variables
  * improves multi-stomp