[FEA] Leverage shared memory with array buffer for writing data to global memory

[ttnghia]

In many operations, we need to write many discrete bytes into global memory. Such writes are not coherent thus it is very inefficient.

We can do better by using shared memory as the scratch pad for such discrete writing, only flush data from there to global memory when the buffer is full. Flushing data to global memory can be done by a single memcpy operation.