Make `arrow-csv` and `arrow-json` optional for applicable crates

### Is your feature request related to a problem or challenge?

Currently importing crates like `datafusion-expr` or `datafusion-common` pulls in `arrow` with its default features, which includes the rather heavy arrow parts `arrow-csv` and `arrow-json`. Since these parts are really only required during physical execution of the respective operators & data sources, that seems rather unnessary.

### Describe the solution you'd like

I see two options:

- set `default-features = false` for the workspace-level `arrow` dependency and then request the right features in the `Cargo.toml` files of the respective datafusion crates
- **better:** do NOT depend on the `arrow` umbrella/meta crate at all but directly on the subcrates that we need

### Describe alternatives you've considered

Keeping the dependencies simple but instead suffer from longer compile times and larger build artifacts.

### Additional context

_No response_

Start by reviewing the workspace-level Arrow dependency and the Cargo.toml files for datafusion-expr, datafusion-common, and the applicable DataFusion crates. Determine which crates require arrow-csv or arrow-json, then verify that unaffected crates no longer pull them in while the relevant builds and tests continue to work.

Written by the indexing model from the issue text.