The actual 13.0.0 DF release uses Arrow 24.0.0, but we need to pick up 25.0.0,
since it brings back the Arrow Schema/Field-to-JSON serialization code (albeit
in a different crate for integration tests).

apache/arrow-rs#2868
apache/arrow-rs#2724

It's now the default HashMap implementation and DF's
planner uses it as well, so we can use std::HashMap everywhere.

Arrow file hash changes and minor changes in the query plan output

Make the `Update`/`Delete` nodes expose `inputs` and `expressions` in order to
let the DF query optimizer work on the `WHERE ...` / `SET col = expr`
expressions. This is slightly hacky:

  - as an "input", we return a `TableScan` node that we don't use after that
    (this is just so that the optimizer knows the input schema for all the
    expressions)
  - return the expressions used by the node and add code to pack/unpack them
    into a list

The point of this is to let DataFusion run the `TypeCoercion` optimization,
without which something like `WHERE float_col > 42` will raise an error (as
after DF 13 these type coercions got removed from other places and moved into
optimizations)

(NB this doesn't work yet, we still get type coercion errors)

(normally it's run only by DataFusion's `create_physical_plan`, but we don't run
that, so we have to execute it manually to get auto type coercion working)

Include `SET` expressions and the predicate if it exists to aid debugging.

These expressions are similar to what DataFusion uses in the `Filter` node and
not doing this seems to break partition pruning (perhaps it stops at the
`Alias` node and doesn't prone anything, didn't investigate in depth).

Copy the `ExprRewriter` visitor from
https://github.com/apache/arrow-datafusion/blob/c50573939d21de40e591c04915d41f7c46a51d0d/datafusion/expr/src/utils.rs#L384-L428
and adapt it to remove aliases from all expressions that the query optimizer
gives back to `Update`/`Delete` nodes.

Make sure the constants are correctly cast and let us detect changes to the
optimizer faster with new DF updates.