Avoid unnecessary null buffer construction when converting arrays to a different type

[etseidl]

Is your feature request related to a problem or challenge? Please describe what you are trying to do.
Related to #6219

While investigating #6219, I found several locations where transforming an array of type T to type U uses the pattern

let array_u = array_t
    .iter()
    .map(|o| o.map(|v| v as U))
    .collect::<PrimitiveArray<U>>()

The initial iter() call emits Option<T> (with nulls becoming None), which is then mapped to Option<U>, with the iter then passed to from_iter()

arrow-rs/arrow-array/src/array/primitive_array.rs

Line 1318 in a693f0f

fn from_iter<I: IntoIterator<Item = Ptr>>(iter: I) -> Self {

which collects the Option<U>s, unwrapping them and collecting into a Buffer, and rebuilding the original null buffer one value at a time.

Describe the solution you'd like
We can avoid one map and the null buffer creation by instead mapping Option<T> to U, and cloning the original null buffer. Adding a from_iter_values_with_nulls() function to PrimitiveArray allows the above to become

let iter = array_t
    .iter()
    .map(|o| match o {
        Some(v) => v as U,
        None => U::default(),
    });
PrimitiveArray<U>::from_iter_values_with_nulls(iter, array_t.nulls().cloned())

This results in a pretty dramatic performance increase.

Describe alternatives you've considered
An alternative is to modify the Array trait to add some sort of transform function that could consume the null buffer from the original array rather than cloning it. But given the use cases I know of are Parquet related, I don't know if that's the right approach.

Additional context

[alamb]

label_issue.py automatically added labels {'parquet'} from #6244

[alamb]

label_issue.py automatically added labels {'arrow'} from #6244