chore(main): release 0.21.0

[laurigates-release-please]

🤖 I have created a release beep boop

0.21.0 (2026-08-10)

Features

• bench: N-point scaling arms for the residency matrix — three points for the price of two (#205) (cc611de)
• bench: N-point scaling arms, one shared cold encode — three points for the price of two (cc611de), closes #175
• core: rebuild the dataset pipeline — O(batch) residency, bucket fitting, grid mode, parallel decode (#197) (ea5733e)

Bug Fixes

• bench: --dry-run must not stop the service it is only asking about (#208) (836ca4c), closes #175

Performance Improvements

• bench: parallelize the residency matrix's encode phase (~8.7x) (#216) (d169fb4)
• core: enable burn-ndarray multi-threads — ~8.5x on the CPU encode phase (#211) (462dc93), closes #209

---

This PR was generated with Release Please. See documentation.

[laurigates-release-please]

🤖 Created releases:

• v0.21.0

🌻