Don't Derive Serialize/Deserialize Serde Implementations for Schema Types

[tustvold]

Which issue does this PR close?

Related to #2300
Part of #2594
Related to #2711

Rationale for this change

Serde derive is a non-trivial proc macro dependency, and the nature of the orphan rule forces the Serialize/Deserialize implementations to be defined in the same crate. This in turn forces json to leak into an arrow-schema crate, see #2711.

It turns out we aren't actually using the generated code, although downstreams potentially are.

What changes are included in this PR?

Removes the serde derives, ensuring we only provide a single way to serialize a schema to JSON, and potentially allowing this to live in a separate arrow-json crate (#2594)

Are there any user-facing changes?

Yes, theoretically users may have been using these serialize implementations. I'm not sure this duplication was intentional.

[tustvold]

Ok I'm rather confused by this, DataType::from doesn't match the format of DataType::to_json for nested types. Perhaps we ought to remove the to_json, etc... methods instead 🤔

Edit: It would appear these relate to the integration test plumbing - https://github.com/apache/arrow/blob/0b67673d045bf6e3523277f6802be126661e3001/docs/source/format/Integration.rst#json-test-data-format

[tustvold]

Closing in favor of #2724