Allow configuring a shorter context to allow a shorter completion.

[jkomoros]

The completion API sets a limit on the length of (input + output). The longer the context, the shorter the completion must be.

There's a balance between passing a lot of context to summarize and getting a long result.

It should be possible to configure that split in ask.

Originally noticed while thinking about #14

[jkomoros]

• [ ] Allow setting answer_length in sample.client.
• [ ] Allow setting prompt_length in sample.client (sometimes you don't want to use the entire prompt; it's more expensive)

[dalmaer]

Consider tiktoken to measure token length.