Introduce observable_sealed_ptr, improve memory footprint in C++17

[cschreib]

Main changes:

• observable_unique_ptr will now optimize memory usage for empty (stateless) deleters for all versions of C++, not just C++20.
• observable_unique_ptr has been split into observable_unique_ptr and observable_sealed_ptr. The former preserves the original API, except that allocation optimisation in make_observable_unique() has been disabled (it prevents writing a safe release() function). The latter has the same API except it is missing void reset(T*) and T* release(); once a raw pointer is acquired, it is "sealed" inside the smart pointer. This enables the allocation optimisation in make_observable_sealed(). You can choose between the two pointers depending on your needs:
  • need to be able to release()? use observable_unique_ptr.
  • no need to be able to release()? use observable_sealed_ptr (reset(T*) can be achieved by assignment).
• Added tests and official support for WebAssembly / Emscripten.

Additional changes:

• Fixed observer_ptr not assignable from observable_unique_ptr with a custom deleter, and not assignable from other observer pointer of same type.
• Fixed control block allocated and not freed when constructing observable_unique_ptr with a T* equal to nullptr (but not directly nullptr).
• Fixed release() from a make_observable_unique() pointer leaving observer pointer with a dangling control block when the user deletes the released pointer.
• Added speed benchmarks.