Log forward intermediates hashes w/pp vs w/o pp

[xmfan]

Stacked PRs:

• ->Log forward intermediates hashes w/pp vs w/o pp #246
• Custom opify triton kernel until local_map functionalization is fixed #245
• Log weight hashes for DSv3 w/ pp vs w/o pp #240

---

Log forward intermediates hashes w/pp vs w/o pp

Intended usage:

> torchrun --nproc-per-node=8 examples/example_ds3_pp.py --rng-seed=42; torchrun --nproc-per-node=4 examples/example_ds3_local_map.py --rng-seed 42
> diff out/0/fw_out.log out/1/fw_out.log 
--- out/0/fw_out.log    2025-11-11 21:00:06.005435737 -0800
+++ out/1/fw_out.log    2025-11-11 21:01:18.998207745 -0800
@@ -1,17 +1,17 @@
 hash=4581379079919370240, norm=214698.34375
-hash=18352098112790593536, norm=2016.0
-hash=9269111720570257408, norm=2016.0
-hash=9146071971375611904, norm=2016.0
-hash=9276605991825178624, norm=2016.0
-hash=9142025768585396224, norm=2016.0
-hash=18384115891391430656, norm=2016.0
-hash=18355862840604098560, norm=2016.0
-hash=9360063322419888128, norm=2016.0
-hash=18320150702933934080, norm=2016.0
-hash=9209896422344753152, norm=2016.0
-hash=21075438881210368, norm=2016.0
-hash=18326941286747078656, norm=2024.0
-hash=9134179653609586688, norm=2024.0
-hash=18431966637432242176, norm=2024.0
-hash=18356249868697075712, norm=2016.0
-hash=9121302173425074176, norm=2024.0
+hash=80466658967158784, norm=2016.0
+hash=9326427062702964736, norm=2016.0
+hash=9192198683184070656, norm=2016.0
+hash=9094632419381739520, norm=2016.0
+hash=9277907813592465408, norm=2016.0
+hash=13440430137933824, norm=2016.0
+hash=72902018968059904, norm=2016.0
+hash=18440621992966094848, norm=2016.0
+hash=9148605246166007808, norm=2016.0
+hash=6157265115545600, norm=2016.0
+hash=9358691131908423680, norm=2016.0
+hash=9360731825489575936, norm=2024.0
+hash=9127494622912708608, norm=2024.0
+hash=18348438938093355008, norm=2016.0
+hash=9250604740851531776, norm=2016.0
+hash=18346538982000558080, norm=2016.0

Currently, fw ins are the same, but the forward is being ran with different rng state between the two setups so there's some numerical differences