chore: refactor out ContextWindow and ChatLogs from the Chat class

[dOrgJelli]

The Chat class was not very easy to use throughout the code. In order to:

• make the implementation simpler
• simplify its construction
• separate sub-concerns

I've created the ContextWindow and ChatLogs classes, which simply contain logic that was previously bundled into the Chat class.

Usage:

// Simple Chat (no context window limit)
const chat = new Chat(tokenizer)

chat.add(...)

// Chat w/ ContextWindow (summarizable)
const llm = new LlmApi()
const contextWindow = new ContextWindow(llm)
const chat = new Chat(tokenizer, llm)

chat.fitToContextWindow()

Motivation: I've been wanting to polish the Chat typings for some time as I think the current implementation does not separate concerns well. This leads to problems like this, where we've regressed and are no longer summarizing the toSummarize messages here https://github.com/polywrap/evo.ninja/blob/962da61051760aad97a2b93d909ccdc9310c9f34/packages/agent-utils/src/llm/Chat.ts#L212

[nerfZael]

I don't think the comment strictly applies now, we could delete it?

[dOrgJelli]

No it's still accurate. If a context window is present, we want to give it a chance to "chunk" a message that is above it's desired size.

[nerfZael]

I'm being a nit, so feel free to ignore, but the message refers to "if the message is larger..." and the functionality (this._contextWindow?.shouldChunk(tokens)) { has no concept of "larger" (it's an internal implementation detail)

[dOrgJelli]

Ah okay, I see what you mean. So a better comment would be something like "See if the context window needs to chunk the message"

[nerfZael]

Yup, exactly.
Or just removing the comment, since it's already evident from the code.
For the most part (99%) I really only believe in comments that explain "why" the code does what it does instead of "what" the code does (cause the latter should always be achievable through cleaner code)