Improve ergonomics / efficiency of `GeneratedContent` coding

[mattt]

As described by @belozierov in #98 (comment):

I receive the LLM response as JSON (decoded using Decodable), which contains the generated Content (decoded via AnyCodable) along with additional metadata. I then: 1. Re-encode the AnyCodable content back into Data 2. Convert that Data into a String 3. Pass it to GeneratedContent.init(json: String), where it is converted back into Data again

To reduce unnecessary transformations (allocations/copies), it would be useful to add GeneratedContent.init(json: Data). It would be even better if GeneratedContent supported Codable directly for JSON data (initially this was a bit confusing, because I assumed this was already how GeneratedContent: Codable worked).

[christopherkarani]

This is a good ergonomics detail. We just added the equivalent path in Conduit, which is a Swift LLM abstraction with a single public import Conduit surface: https://github.com/christopherkarani/Conduit

The branch adds GeneratedContent(json: Data) plus jsonData so callers can round-trip provider JSON data without unnecessary string conversion: https://github.com/christopherkarani/Conduit/tree/codex/conduit-foundationmodels-provider-expansion