Implement a more accurate hardware model.

[miladHakimi]

The current model changes Enzyme's tape loads and stores with Tloads and Tstores. It makes the all the instructions after the barrier be dependent on it and models the memory latency by using a delay. It does not support double buffering right now and the delay might not be fixed in different benchmarks. Moreover, using the noop native operations in salam we want to add the comparison between the recompute and store-all models.

• Implement actual writes and read request to memory for the barriers.
• Implement double buffering.
• Implement the store-all execution model.