when_all returns pointer of futures

[kyungjoo-kim]

@hcedwar I put an issue after the discussion about packing a group of futures.

/// currently
struct block {
  int m, n;
  value_type *ptr;
  future_type future;
};

view<block**> blocks; 

doSomethingWithBlocks(blocks); // blocks(i,j).future has a handle for task.

// the whole blocks need to be dependent on another task

view<future_type*> deps("deps", blocks.m * blocks.n);
int k=0;
for (int i=0,k=0;i<blocks.m;++i)
  for (int j=0;j<blocks.n;++j,++k) 
     deps[k] = blocks(i,j).future;

Kokkos::respawn( this, when_all(blocks.m*blocks.n, deps), Priority::Regular);
// this depedence mechanism do multipl copy of futures (internally copy the array again)
----------
// the following may reduce the overhead in this situation
auto deps = when_all(blocks.m*block);
int k=0;
for (int i=0,k=0;i<blocks.m;++i)
  for (int j=0;j<blocks.n;++j,++k) 
     deps[k] = blocks(i,j).future;

Kokkos:respawn( this, deps, Priority::Regual );

[hcedwar]

A more robust interface:

  when_all( blocks.m * blocks.n , [&]( int i ) { return blocks(i/blocks.n,i%blocks.n).future ; } );