Count spend tokens on issue context (all tokens spent mean issue is too complex and is not solvable by AutoPR)

[Konard]

It also should not begin solving the issue if the issue itself uses too many tokens even after summarization. AutoPR should always resummarize the issue context between steps and keep track of how much tokens left.

Some context may be temporary offloaded to pull request comments if it is not required for the next step.

This may also require doing one change at a time, and after that change is done it can be removed from context.

Related to #88 and #93

[irgolic]

I might be misunderstanding, but this isn't a viable way of using the LLM. If you ask it to summarize something, it'll do it, into whatever short form you like.

Another way to interpret what you're saying is that it should split issues if they're too long. I'm down to explore this, as a separate trigger on the issue itself (akin to #86 (comment)).