Simplify truncation logic to only keep the last message and exclude embeddings

[nicohrubec]

• Instead of keeping as many messages as fit into our limit, we now discard everything but the newest message. So the result is an array of messages with one item. This remaining messages optionally gets truncated if it exceeds the byte limit.
• So far embeddings went through the same truncation logic as other messages. From now on embeddings should be set as gen_ai.embeddings.input and not be truncated.

[linear]

JS-1521

[github-actions]

A PR closing this issue has just been released 🚀

This issue was referenced by PR #18906, which was included in the 10.37.0 release.