Which issue does this PR close?

• Closes HashJoinExec / NestedLoopJoinExec projection Some(vec![]) becomes None after ser/de #23083

Rationale for this change

HashJoinExec and NestedLoopJoinExec carry projection: Option<Vec<usize>> but the proto field is repeated uint32. Proto3 can't tell None from Some(vec![]), and the decoder treats both as None. Some(vec![]) is reachable in real plans — try_embed_projection produces it for SELECT count(1) … JOIN … (#20191) — and after a round-trip the join silently switches from "emit zero columns" to "emit all columns".

FilterExec has a workaround for the same limitation; these two execs were missed.

What changes are included in this PR?

Encode Some(vec![]) as the single-element sentinel [u32::MAX] (never a valid column index); recognise it on decode. Everything else goes through unchanged.

// encode
projection: match exec.projection.as_ref() {
    None => Vec::new(),
    Some(v) if v.is_empty() => vec![u32::MAX],
    Some(v) => v.iter().map(|x| *x as u32).collect(),
},

// decode
let projection = match hashjoin.projection.as_slice() {
    [] => None,
    [u32::MAX] => Some(Vec::new()),
    indices => Some(indices.iter().map(|i| *i as usize).collect()),
};

Applied symmetrically to HashJoinExec and NestedLoopJoinExec.

Are these changes tested?

yes, add roundtrip test case

Are there any user-facing changes?