fix(rust): make 100 * pl.col(pl.Boolean).mean() work

[Wainberg]

Closes #7651 based on the implementation suggested by @advoet here.

Before:

pl.DataFrame({"a": [True, True, True, False, False]}).select(100 * pl.col("a").mean())
shape: (1, 1)
┌─────────┐
│ literal │
│ ---     │
│ i32     │
╞═════════╡
│ 0       │
└─────────┘

After:

>>> pl.DataFrame({"a": [True, True, True, False, False]}).select(100 * pl.col("a").mean())
shape: (1, 1)
┌─────────┐
│ literal │
│ ---     │
│ f64     │
╞═════════╡
│ 60.0    │
└─────────┘

[ritchie46]

This branch already happens in the float_type call?

Edit. I see that there are two float_type functions that are inconsistent. Boolean should be added to the float_type function, not as an exclusion here.

[Wainberg]

I'll be honest, I just blindly copied @advoet's code from the linked issue. But it does fix the bug.

[ritchie46]

It does fix it, but it is a bandaid.

[Wainberg]

Happy to implement a more robust fix, just let me know where the "right" place to change would be. It's an important fix because people like to use 100 * boolean_column.mean() to calculate percentages.

[mcrumiller]

There are a lot of cases where boolean is treated as a u8 in these types of operations. It might be worth adding is_numeric_or_bool() for those many cases scattered throughout the codebase (perhaps different PR).

[mcrumiller]

Eh, I just perused through and it feels like it's not worth it. Go ahead with this one.

[Wainberg]

I'm just not familiar enough with the Rust codebase to know what a non-bandaid fix would look like.

[mcrumiller]

He just means at the very top of the file do this instead of your else clause below:

fn float_type(field: &mut Field) {
    if (field.dtype.is_numeric() || field.dtype == DataType::Boolean)
        && field.dtype != DataType::Float32
    {
        field.coerce(DataType::Float64)
    }
}

[Wainberg]

ohhh I didn't see Ritchie's edit to his original post! thanks :)

[mcrumiller]

He may have been noting the other float_type function in expr_dyn_fn.rs but this is actually never used.

[Wainberg]

@ritchie46 updated this PR based on your suggestion:

Edit. I see that there are two float_type functions that are inconsistent. Boolean should be added to the float_type function, not as an exclusion here.