Faster subsetting for cohorts (#397)

* Faster subsetting for cohorts

Closes #396

* tpying

Fix default int on windows, numpy<2 (#395)

* Avoid rechunking when preferred_method="blockwise"

* Add new numpy1 environment

* try int_ instead of intp

* Use uintp instead

* Use np.uint instead

* more fixes

* Add test

* fix

* fix again

Preserve dtype better when specified. (#389)

* Preserve dtype better when specified.

* Add one more test

* tweak test

* more test

* [revert] test with Xarray PR branch

* tweak

* show versions

* Drop python 3.9, use ruff

* switch to Ruff

* fix mypy

* remove toctrees

* fix

* one more

Expand groupby_reduce property tests (#385)

* Expand groupby_reduce property tests

* Add back var, std

* cast quantile result

* Revert "Add back var, std"

This reverts commit 805b8d3.

* pin numpy in benchmark env

* Add benchmarks as test

Optimize for-loop merging of cohorts. (#378)

* Optimize for-loop merging of cohorts.

Do this by skipping perfect cohorts that we already know about.

* Add new benchmark

* Fix

* Cleanup print statements

* minimize diff

* cleanup

* Update snapshot

Cubed blockwise (#357)

* Initial minimal working Cubed example for "blockwise"

* Update minimum cubed version that includes cubed-dev/cubed#448

* Fix mypy errors

* Update documentation with a 'blockwise' example for Cubed

import `normalize_axis_index` from `numpy.lib` on `numpy>=2` (#364)

* import `normalize_axis_index` from `numpy.lib` on `numpy>=2`

* import the right thing

`method` heuristics: Avoid dot product as much as possible (#347)

* Another `method` detection optimization

* fix

* silence warnings

* silence one more warning

* Even better shortcut

* Update docs

Fix nanlen with strings (#344)

* Fix nanlen with strings

Closes pydata/xarray#8853

* fix windows

* Silence warnings

Fix direct quantile reduction (#343)