StructArray::from panics when passed empty vector

[bjchambers]

Describe the bug
StructArray::from(vec![]) causes a panic because it tries to index into the list of arrays to check their lengths here https://github.com/apache/arrow-rs/blob/master/arrow/src/array/array_struct.rs#L212.

To Reproduce
Steps to reproduce the behavior:

Expected behavior
A clear and concise description of what you expected to happen.

Additional context
Add any other context about the problem here.

[bjchambers]

I guess perhaps this is reasonable, since without their being at least one array it would be impossible to know the length of the struct array to create.

[tustvold]

I think having it just create a 0 length array would be sensible and avoid surprise panics. There are niche use-cases where a non-zero row count may be desirable, see #1536, but I think we can deal with that as needed?

[bjchambers]

I'm also seeing this with new_null_array(&DataType::Struct(vec![]), N). I wonder if the best path would be the following:

1. Introduce a method on StructArray fn new(data_type: &DataType, length: usize, fields: Vec<(Field, ArrayRef)>) -> Self which just checks that all the field arrays have the given length. This could also have a variant that takes the null Buffer.
2. The existing StructArray::from(...) methods use that based off the length of the first field (throwing a reasonable error when called with 0-fields.
3. The new_null_array can then use that method rather to ensure that it works even on zero length arrays.

Then users who have to handle the non-zero length array of 0 fields would be able to call StructArray::new(...) explicitly.

[tustvold]

I think you could get away with a simpler signature, such as

StructArray::new_with_len(fields: Vec<(Field, ArrayRef)>, len: usize) -> Self

But yes that makes sense to me, I hadn't thought of the null use-case

[bjchambers]

I think the issue with just creating it of length 0 is that if you then go to put that into a record batch or enclosing struct array of a given length you'll have problems because the lengths don't line up. So, it seems like general idea of having a constructor from an explicit length makes sense, and then the various from methods can use that.

[tustvold]

I think this is closed by #4064