Dask-CuDF: use default Dask Dataframe optimizer

[madsbk]

In order to use the new HighLevelGraph optimization work in Dask/Distributed, this PR makes dask_cudf.Dataframes use the default Dask optimizer.
Previously, we have been explicitly materialized the HighLevelGraphs when calling submit() and compute() on dask_cudf.Dataframes.

Overall, this should improve performance but by default low-level task optimizations are disabled, which might have a negative impact. High-level optimizations are done in any case and we are working on moving all low-level optimization to high-level but currently low-level optimization such as array slicing is only supported by the low-level.

I don't think we will be missing any low-level optimizations related to Dataframes so I think we should follow Dask on this one and disable low-level optimizations by default.
It is possible to enable low-level optimizations explicitly by setting the Dask config like:

dask.config.set({"optimization.fuse.active": True})

cc. @jakirkham, @quasiben, @beckernick, @VibhuJawa

[jakirkham]

cc @rjzamora

[jakirkham]

Interestingly this optimize function has been around since Jim initial setup Dask-cuDF (called Dask-GDF at the time). Am curious why it was needed. @jcrist do you happen to remember? If not, no worries 🙂

[madsbk]

rerun tests

[codecov]

Codecov Report

❗ No coverage uploaded for pull request base (branch-21.08@aa67b13). Click here to learn what that means.
The diff coverage is n/a.

❗ Current head 75194d7 differs from pull request most recent head 3fc25c8. Consider uploading reports for the commit 3fc25c8 to get more accurate results

@@               Coverage Diff               @@
##             branch-21.08    #8581   +/-   ##
===============================================
  Coverage                ?   82.61%           
===============================================
  Files                   ?      109           
  Lines                   ?    17850           
  Branches                ?        0           
===============================================
  Hits                    ?    14747           
  Misses                  ?     3103           
  Partials                ?        0           

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update aa67b13...3fc25c8. Read the comment docs.

[jakirkham]

Was wondering if we should be setting this somehow to ensure it is 1. However that appears to be the default anyways. So that doesn't seem needed

The other culling steps happen naturally as part of optimize anyways so no need to reproduce them otherwise either

[jakirkham]

@gpucibot merge

[jakirkham]

Thanks Mads! 😄