Hash UnionArrays

[friendlymatthew]

Which issue does this PR close?

• Closes Make UnionArrays hashable #18717

Rationale for this change

This PR adds hash support for Union data types, enabling group by, distinct, hash joins, and aggregations on union-typed columns

hash_union_array hashes each child array once. Then for every row, it uses the type id and offset to retrieve the appropriate hash value

[friendlymatthew]

I found that clippy's suggestion hurt readability

error: the loop variable `i` is used to index `hashes_buffer`
   --> datafusion/common/src/hash_utils.rs:351:14
    |
351 |     for i in 0..array.len() {
    |              ^^^^^^^^^^^^^^
    |
    = help: for further information visit https://rust-lang.github.io/rust-clippy/rust-1.91.0/index.html#needless_range_loop
    = note: `-D clippy::needless-range-loop` implied by `-D warnings`
    = help: to override `-D warnings` add `#[allow(clippy::needless_range_loop)]`
help: consider using an iterator and enumerate()
    |
351 -     for i in 0..array.len() {
351 +     for (i, <item>) in hashes_buffer.iter_mut().enumerate().take(array.len()) {
    |

[adriangb]

is 128 the max? I feel like just using a HashMap and then doing a get() below instead would be more readable and the same performance (maybe unless we 128 single item arrays or something...)

[friendlymatthew]

Yes. Type ids are i8s, and negative values should not be valid so we bound the total number of spaces to 128

In terms of using a direct-indexed Vec by type id over a hashmap, I just chose not to overcomplicate things. Fwiw, I've yet to write benchmarks for this (I'll plan on doing that later on) and if this part shows up in a profile, I'll happily update it

[friendlymatthew]

Scratch that. I'll use a hashmap