KREA2 and other models is significantly slower comparing to earlier builds.

### Custom Node Testing

- [x] I have tried disabling custom nodes and the issue persists (see [how to disable custom nodes](https://docs.comfy.org/troubleshooting/custom-node-issues#step-1%3A-test-with-all-custom-nodes-disabled) if you need help)

### Expected Behavior

Seriously guys, what on earth are you doing with memory and speed optimisations.
All builds later than 16.4 get slower and slower. Memory management (pinned memory) gets worse almost every new build. Users have to disable all optimisations if possible (what's the point of adding such optimisations?? )

default start args. + --use-sage-atention, fp8_matrix_mult
windows-standalone
5060 16GB, 32RAM
all models are on a fast HDD the system and pagefile are on an SSD
cu128

0.16.4 taken as a base
pinned memory, model management, ops, quant ops, model base....

merged all missing models and some other stuff from 26.1 and 27 and 30
I had to add int8_tensorwise and convrot_w4a4 support from "scratch" but it does not matter

by default 23,26,27 builds were bad with KREA 2 and LTX (26.1 was the best among 3)
they all eventually started reloaded all models every new generation even for just new seeds.

0.16.4 is the best for LTX but it does not support KREA 2 and others...

that's why I had to start merging.

the merged build runs fine LTX and KREA 2, it runs MINIMAX H3 but it does not listen to prompts atm...
I moved to 0.30.1 it was working OK with KREA2 and MINIMAX, until 0.33.1.and even worse in 0.34.0

For example KREA2 speed.

my merged build 0.16.4 + 26.1 + 27 + 30:
first run 183s
all others 39-43 sec, depending on lora and prompt changes
never slows down.

0.30.1
1st around 250s
all others 60-90s, can slow down

0.33.1
first run 310s
second 210s
all others around 110s

0.34.0
1st run 360s
all others 120s +

same workflow, same seed...

### Actual Behavior

it should run 3 times faster than in the latest build

### Steps to Reproduce

generate a picture several times with different speed

### Debug Logs

```powershell
not needed
```

### Other

_No response_

No source file, test, or entry point is named. Reproduce the same workflow across the listed ComfyUI builds with identical settings, then collect profiling or debug evidence around model reloads and pinned-memory behavior; done means identifying the regression and showing that repeated generations no longer slow down.

Written by the indexing model from the issue text.