- Also renamed the badly named `Get_global` to `Access`.

- Initialization now needs to be handled via running the corresponding code explicitly. In particular `Tensor.init_params` will run the forward code of tensors from the `params` field.

- Virtual nodes and inlining now also work across routines. This required changing the API to pass the `optimize_ctx` optimization context.

To avoid misleading behavior of `device_to_device` data movement, non-constant materialized tensor nodes are represented in contexts making use of them, even when the underlying array is on host. This way the logic remains the same regardless of whether a backend shares memory with the host.