Simplify Schema implementation in Rust

We have a custom schema implementation, but we so often convert it to/from Arrow schemas, it might make sense to implement it as:

pub struct Schema {
    schema: Arc<arrow_schema::Schema>,
    /// Field ids, in depth-first order
    field_ids: Vec<i32>,
}

This would be quick to convert to Arrow when we need to.

Getting by field id wouldn't be slower, as we already do a linear search of fields anyways. We could make it fast by doing field_ids: HashMap<i32, Arc<arrow_schema::Field> if we wanted to.

Getting the id of a given field would be harder, though.

Start by locating the current custom Schema implementation and every path that converts it to or from Arrow schemas. Review how field-id lookup and retrieving an id for a field currently work. Done means the representation can use an Arrow schema without slowing or breaking those behaviors, with tests covering the conversion and lookups.