Streamlined representation of PendingCounts data structure.

(1) We now allocate a contiguous array of memory to hold count information
for each node in a frame, allocating either 8 bits to hold a
PendingCounts::PackedCounts structure (if the dead and pending count
fields can fit in 8 bits), or an aligned region of 64 bits to hold a
PendingCounts::LargeCounts structure (if the counts are too large to
fit in the 8-bit representation).  The byte offset and representation
of the data for a given node are now held in a 32-bit handle.  This
makes all the count data contiguous for a given frame (better cache
locality), and avoids the need for the FlatMap that we were using to
hold a mapping from id to LargeCounts for those nodes that needed
the overflow space.

(2) Sped up the fast path of adjustments needed when normal non-merge
nodes complete by adding a specialized "adjust_for_activation" routine
that does all the necessary bookkeeping in one routine, rather than
having to manipulate the count data structure several times.

(3) Several tensorflow/core/common_runtime/executor.cc changes related to this:
. Stored the new PendingCounts::Handle for a given node in the ItemState
data structure, rather than in a separate frame_local_ids_ array.  This
means that we touch one less cache line per edge in ActivateNodes.

. Changed initialization code to use the new PendingCounts::AllocateHandle
routine for allocating space to hold the counts for a node.  This allowed
simplifying the ControlFlowInfo that is computed at initialization (we
just need the unique set of frame names now, rather than a mapping from
frame name to number of nodes in that frame).

(4) Updated tensorflow/core/common_runtime/pending_counts_test.cc to reflect
the new interface and to test the new compound adjust_for_activation routine.

Speeds up an image processing benchmark on my desktop from 3370 images/sec
to 3540 images/sec (+5.0%).
Change: 146684811