Change the DAG to have separate nodes for operations and arrays

[tomwhite]

Currently there is a one-to-one correspondence between operations and arrays: one operation produces one array, and is represented by one node in the DAG.

However, this won't be true when we support multiple outputs (#69) where one operation can produce multiple arrays.

Similarly, it also won't support optimizations like sibling fusion (in Apache Beam, and from FlumeJava originally), where operations that have the same inputs and produce different outputs can be fused into one operation producing multiple outputs. (Wukong calls this "task clustering".)

So we need to change the DAG representation so that one operation can produce multiple arrays, and this means breaking nodes into two types: operations and arrays.

This PR does exactly that. An example DAG (from test_add_with_broadcast) in the current representation

now looks like this:

Note:

• Operations are shown as boxes with rounded corners, arrays are rectangles.
• Colours indicate the following: pink for a blockwise operation, green for a rechunk operation, orange for a materialized array, white for an operation that is not run in its own pipeline or for an array that is not materialized to disk as a Zarr array ("virtual arrays")

I wondered about having an option to visualize DAGs as we currently do, but that won't work when there are multiple outputs, so it may be best just to have the more general representation.

[tomwhite]

@TomNicholas any thoughts on this one before merging?

[TomNicholas]

This looks great! I also asked my colorblind friend if he could distinguish all the colors and he gave it a thumbs up 👍

[tomwhite]

I also asked my colorblind friend if he could distinguish all the colors and he gave it a thumbs up 👍

Thank you!