…enter segments in a way that seems more correct (or at least more self-consistent) to a native Chinese speaker (#613)

See discussion at #539 (comment) for explanation

* Add `sideEffects: false` to `libesm/package.json`

* Make tweaks I proposed in convo at #612

* Fix build for Unix (and break it again for Windows in the process)

* Use `node --eval`

* --target es6 for ESM build

@fisker points out at #612 (comment) that ESM modules aren't usable by ES5 environments anyway, so targeting ES5 is pointless

---------

Co-authored-by: Mark Amery <markrobertamery@gmail.com>

Somehow I screwed up #613 and merged it with the tests failing and with the sentence I was actually using in the test inconsistent with the one I claimed to be using in the comment above. Also, even if I'd got it right, I wouldn't've actually avoided hitting the inconsistency in Intl.Segmenter's tokenization rules that that PR was specifically trying to avoid, because it considers 他有 (he has) to be one word; I should've used 她有 (she has) which the segmenter sees as two words. This fixes both mistakes.

* Fix diffSentences for old Safari versions

Resolves #615

* Add release notes