Refactor `mean()` reducer to use `FExpr`

[samukweku]

• refactor mean() reducer to use FExpr;
• fix median() to use float64 for void columns.

WIP for #2562

[samukweku]

@oleksiyskononenko I do not understand the rules for void when computing the mean. If you have some time, I'd love your feedback on it

[oleksiyskononenko]

@samukweku So what is your question on voids?

[oleksiyskononenko]

@samukweku I see your problem now. I guess our existing mean() implementation is not consistent when it comes to groupby context. For void columns

• with no groupby the column's return type is SType::VOID: https://github.com/h2oai/datatable/blob/main/src/core/expr/head_reduce_unary.cc#L223
• in the groupby context it is SType::FLOAT64: https://github.com/h2oai/datatable/blob/main/src/core/expr/head_reduce_unary.cc#L245-L246

On this PR we have already fixed that inconsistency and now only the failing tests have to be fixed. Go ahead and adjust the types in the reference frames to dt.float64.

[oleksiyskononenko]

Is there anything else you're planning to do on this PR? Not sure why the build is failing.

[samukweku]

@oleksiyskononenko Just ur review and merge is what's left

[oleksiyskononenko]

The build is read, so we first need to make it green before merging. Can you re-trigger the build to see if it was a temp glitch on AppVeyor?

[samukweku]

Sure, will do

[oleksiyskononenko]

@samukweku once I merge https://github.com/h2oai/datatable/pull/3412, please remove your changes to test_median* tests and merge main onto your PR. test_median* tests have typos and testing mean() instead of median().

[oleksiyskononenko]

Nevermind, I've already made the required fixes. See if you have anything else to be done for this PR, otherwise, I guess it is all good to be merged.

[samukweku]

Thanks @oleksiyskononenko all good from me. Ok to merge