py-1.21.0

## 🚀 Performance improvements

- Use BitmapBuilder in yet more places (#20868)
- Make an owned version of append (#20800)
- Use BitmapBuilder in a lot more places (#20776)

## ✨ Enhancements

- Stabilize methods/functions (#20850)
- Add `linear_space` (#20678)
- Improve string → temporal parsing in `read_excel` and `read_ods` (#20845)
- Implement df.unique() on new-streaming engine (#20875)
- Experimental credential provider support for Delta read/scan/write (#20842)
- Allow column expressions in DataFrame `unnest` (#20846)
- Auto-initialize Python credential providers in more cases (#20843)
- Add unique operations for Decimal dtype (#20855)
- Add NDJson sink for the new streaming engine (#20805)
- Support nested keys in window functions (#20837)
- Add CSV sink for the new streaming engine (#20804)
- Periodically check python signals ('CTRL-C' handling) (#20826)
- Experimental unity catalog client (#20798)
- Support cumulative aggregations for `Decimal` dtype (#20802)
- Account for SurrealDB Python API updates (handle both `SurrealDB` and `AsyncSurrealDB` classes) in `read_database` (#20799)
- Drop `nest-asyncio` in favor of custom logic (#20793)
- Improve window function caching strategy (#20791)
- Support `lakefs://` URI for delta scanner (#20757)
- Additional support for loading `numpy.float16` values (as Float32) (#20769)

## 🐞 Bug fixes

- Warn if asof keys not sorted (#20887)
- Ensure explicit values  given to `column_widths` override autofit in `write_excel` (#20893)
- Avoid name collisions and panicking in object conversion (#20890)
- Incorrect scale used in `log` and `exp` for Decimal type (#20888)
- Don't deep clone manuallydrop in GroupsPosition (#20886)
- Fix DuplicateError when selecting columns after `join_where` or cross join + filter (#20865)
- Incorrect `Decimal` value for `fill_null(strategy="one")` (#20844)
- Fix one edge case (out of many) of int128 literals not working (#20830)
- Add height check to frame-level row indexing when key is int (#20778)
- Remove `assert` that panics on `group_by` followed by `head(n)`, where `n` is larger then the frame height (#20819)
- Selectors should raise on `+` between themselves (#20825)
- Fix panic `InvalidHeaderValue` scanning from S3 on Windows (#20820)
- Fix `clip` for `Decimal` returning wrong values (#20814)
- Incorrect height from slicing after projecting only the file path column (#20817)
- Shift mask when skipping Bitpacked values in Parquet (#20810)
- Error instead of truncate if length mismatch for several `str` functions (#20781)
- Support cumulative aggregations for `Decimal` dtype (#20802)
- Allow `is_in` values to be given as custom `Collection` (#20801)
- Propagate null instead of panicking in `pl.repeat_by()` (#20787)
- Do not print sensitive information to output on `POLARS_VERBOSE` (#20797)
- Ignore file cache allocation error if `fallocate()` is not permitted (#20796)
- Incorrect logic in `assert_series_equal` for infinities (#20763)

## 📖 Documentation

- Update source URL for `legislators-historical.csv` (#20858)
- Update ML part of ecosystem user guide page (#20596)

## 🛠️ Other improvements

- Disable 'catalog' in build (#20897)
- Implement negative slice for new streaming IPC (#20866)
- Debloat Series bitops (#20873)
- Reduce python map bloat (#20871)
- Remove todo and test restriction for new-streaming (#20861)
- Dispatch to the in-mem engine for `AExpr::Gather` (#20862)
- Dispatch to the in-memory engine for multifile sources (#20860)
- Add tests for open issues (#20857)
- Mark 'register\_startup' as unsafe (#20841)
- Reduce mode bloat (#20839)
- Rename `ContainsMany` to `ContainsAny` (#20785)
- Unpin NumPy in type checking workflow (#20792)
- Add various tests (#20768)
- Small drive-by's (#20772)
- Touch the upload probe for the remote benchmark (#20767)

Thank you to all our contributors for making this release possible!
@alexander-beedie, @arnabanimesh, @braaannigan, @burakemir, @coastalwhite, @etiennebacher, @ion-elgreco, @itamarst, @lukemanley, @mcrumiller, @nameexhaustion, @orlp, @ritchie46 and @stinodego