Icepool

* Removed `Die.keep()`. Use `Die.pool(...).sum()`.
* `highest`/`lowest` returns empty die if any of the input dice are empty.

* Free-function form of `lowest`, `highest` now selects between algorithms for better performance and generality.
* Removed `die.lowest, `die.highest`.
* `die.min_outcome`, `die.max_outcome` no longer take arguments beyond `self`.
* No more (public) `repeat_and_sum`, as this is redundant with `` and `die.keep()`.
* No more public `die.keep_lowest_single()`, `die.keep_highest_single()`.
* `count_dice` arguments are now keyword-only.
* `die.zero()` no longer reduces weights to 1.
* Update PyPi classifiers.

* Removed `min_outcome` parameter from die construction.
* Operations on tuple outcomes are now performed recursively, to match the fact that die expansion on construction is recursive.
* Renamed `Die.reduce()` to `Die.reduce_weights()`.
* Added `reduce()` and `accumulate()` functions analogous to the functions of the same name from functools/itertools.
* Add CSV output.
* Renamed `Die.markdown()` to `Die.format_markdown()`.
* Added `star` parameter to `sub`, `explode`, `reroll`, `reroll_until` methods. If set, this unpacks outcomes before giving them to the supplied function.
* Added experimental `JointEval` class for performing two evals on the same roll of a pool.

* Operators other than `[]` are performed element-wise on tuples.
* Rename `DicePool` to just `Pool`. Merge the old factory function into the constructor.

* Fix denominator_method='reduce' in die creation.
* Fix outcomes consisting of empty tuple `()`.
* `apply()` with no dice produces an empty die.

Retired the `EmptyDie` / `ScalarDie` / `VectorDie` distinction.

* There is now only one `Die` class; it behaves similarly to how `ScalarDie` used to.
* There is no more `ndim`.
* The `[]` operator now forwards to the outcome, acting similar to what `VectorDie.dim[]` used to do.
* Removed `PoolEval.bind_dice()`. It was cute, but I'm not convinced it was worth spending API on.