MultiVectorRetriever does not obey self.chunk_size

[TanGentleman]

Since the process follows like this:

• Separate inputs into pages to make parent docs
• For each page, create a summary/qa list
• Create a list of child_docs from these summaries
• The vector database stores these shortened docs, and aliases them to the parent documents

These parent docs are prioritized to "page" correctly, like for a pdf. This may lead to unresolved context limits or unexpected behavior without better checks on the size of doc.page_content. Since it's the parent document excerpts getting passed to the LLM at the very end, we need to calculate character count and respect the token limits.