Once an object got into the cache c.objCounters.Size() is not zero until it
leaves the cache, even if there are no new objects coming in. This means we're
looping here without any stops while we have objects in cache and a lot of
cycles can be made even with a very small ops/s load (multiplied by the number
of shards). This loop also makes a copy of object list on each iteration which
leads to _a lot_ of allocations and associated GC pressure.

As a minimal fix we can just iterate over the list once per 100ms, 20 threads
can add enough pressure to real HDDs for them to be busy all this time and we
won't be wasting cycles for nothing.

Signed-off-by: Roman Khimov <roman@nspcc.ru>

1. Make it a separate active (timer!) entity.
2. Simplify in-transit checks by maintaining a single map of items, don't
   allocate additional maps.

Signed-off-by: Roman Khimov <roman@nspcc.ru>

Signed-off-by: Roman Khimov <roman@nspcc.ru>

Signed-off-by: Roman Khimov <roman@nspcc.ru>

…3414

This allows to have multiple batches flushing at the same time similar to
multiple objects.

Signed-off-by: Roman Khimov <roman@nspcc.ru>

…Objs

Adding into flushObjs _after_ object submission to worker thread is inherently
wrong, that thread can flush (and attempt to delete from flushObjs) the object
before it's added into flushObjs.

Signed-off-by: Roman Khimov <roman@nspcc.ru>

Doing it per-address is excessive, we will do it anyway for big objects
(since we're waiting on send anyway) and small ones just manage the address
slice (quickly).

Signed-off-by: Roman Khimov <roman@nspcc.ru>

It's not needed.

Signed-off-by: Roman Khimov <roman@nspcc.ru>

1. Sleep more, enough for the new batch to appear on its own (wrt time limit),
   ticker works perfectly for this.
2. Once the time has passed we're always getting a set of current objects and
   sort them by size.
3. This means we already have natural batches and we don't even need to
   allocate slices for them. Timer management becomes irrelevant as well.
4. Push batches into threads and be happy, just don't forget about big objects
   (subslice of one) and flushing the last object from the set.

Signed-off-by: Roman Khimov <roman@nspcc.ru>