Add experimental Dask-on-Mars support

[loopyme]

What do these changes do?

Add experimental Dask-on-Mars support, which make it possible to run Dask collections on Mars.

[qinxuye]

We don't have the habit to use sort of experimental stuff, cuz users anyway need to change their code. I recommend to rename experimental to contrib, like contrib/dask.

[qinxuye]

What's more, you can create a separate CI test in platform-ci, refer to https://github.com/mars-project/mars/blob/master/.github/workflows/platform-ci.yml to add a new ci in which dask installed.

[loopyme]

I am not familiar with CI tests, will this works?

[qinxuye]

Seems dask test timeout. Could you please address the issue please?

[loopyme]

Working on that

[qinxuye]

Is this PR ready for a review? maybe we could mark it as some milestone now?

[qinxuye]

LGTM overall, coverage could be increased a bit.

[hekaisheng]

Will convert_dask_collection method be exposed to users? It seems that the recommended way to execute Dask collections is df.compute(scheduler=mars_scheduler), could you add docs in user guide?

[loopyme]

Will convert_dask_collection method be exposed to users? It seems that the recommended way to execute Dask collections is df.compute(scheduler=mars_scheduler), could you add docs in user guide?

Exposing convert_dask_collection method to user sounds better to me, since it may be more robust than mars_scheduler. Besides, it allows users to modifying the task graph based on the original Dask collections using Remote API or other Mars features. If user only need to run Dask collections, mars_scheduler will be better choice.

I am working on doc

[qinxuye]

Spelling check failed: https://github.com/mars-project/mars/pull/2289/checks?check_run_id=3342458497

[loopyme]

Sorry about the typo

[qinxuye]

LGTM

[wjsi]

LGTM

[qinxuye]

Congrats, look forward to seeing more contributions from you.