This reverts commit 884a96b2.

Allow the load index and result to be fused with other per-element ops.
The graph edge into the values is considered as a "gather" edge and
is not allowed to be fused (we must force values to be written to
memory so we can index into them).

Removed the special kernel for gather loads, not needed any more.

With padding and arithmetic for now (extend padding to allow different
amounts before and after the data).

Non-deterministic for now due to use of atomic float add.