Choose an allocator for memtx in box.cfg

[alyapunov]

Slab allocator, which is used for allocation tuples, has a certain disadvantage - it tends to unresolvable fragmentation on certain workloads (size migration).

Provide an option in box.cfg that allows to choose some other allocator for memtx tuples. At least it must be simple allocator like hpall (there must a branch). At most - log structurea allocator (like lsall, also on the same branch).

There's also an idea to make that option per space, but I'm not sure.

[filonenko-mikhail]

This day I prefer such option per space, to enable it only for spaces with some array data at field. These are spaces with:

• aggregator with sources (two or more tuples from different IT-departments about the same customer makes golden record in mdm systems)
• denormalized foreign keys in field with type array
• any multiindex data: person emails, phones, etc

[EvgenyMekhanik]

After discussion with @lyapunov and @Korablev77, we decided the following:

1. The allocator class should look like this:

   class Allocator {
   public:
        static inline void* alloc(size_t size);
        static inline void  free(void *ptr, size);
        static inline void  create(/* specific for this allocator arguments */);
        static inline void destroy(void);
        static inline void stats(struct allocator_stats *stats);
   private:
        /* specific for allocator private static members */
   };

2. There are several methods for getting allocator statistic:
   • box.slab.stats for small allocator custom statistic
   • box.system.stats for system allocator custom statistic
   • box.slab.info (deprecated) for common allocators statistic
   • box.alloc.info same as previous, but new recommended method.
3. Delayed free mode removed from allocator.