Adding support for diffusiongemma-26B-A4B-it

[AnishPahilajani]

You can run diffusion gemma on 2 spyre cards using the following command

torchrun --nproc-per-node 2 scripts/test_diffusion_gemma.py --model /models/diffusiongemma-26B-A4B-it --tp --max-new-tokens 256 --max-denoising-steps 8

output I get

 blue is due to a phenomenon called **Rayleigh scattering**.

Here is the step-by-step breakdown:

### 1. Sunlight is a rainbow
Although sunlight looks white to us, it is actually made up of all the colors of the rainbow (red, orange, yellow, green, blue, indigo, and violet). Each color travels as a wave with a different wavelength. Red light has long, lazy waves, while blue and violet light have short, choppy waves.

### 2. The atmosphere is an obstacle course
Earth’s atmosphere is filled with gases (mostly nitrogen and oxygen). These gas molecules. 

### 3. Scattering
When sunlight hits the atmosphere, it strikes the gas molecules and gets scattered in all directions. 
*   The **long waves** (red and yellow) pass through the atmosphere relatively undisturbed, hitting very molecules.
*   The **shorter waves** (blue and violet) strike the gas molecules and get scattered (scattered) in every direction.

### 4. Why do we see blue and not violet?
Since the blue light is being scattered in every part of the sky, when you look up, your eyes catch that scattered blue light coming from every part of the atmosphere

--- throughput ---
  max_denoising_steps : 8
  generated tokens    : 256
  wall time           : 266.74s
  throughput          : 1.0 tok/s

NOTE: throughput is dominated by Spyre↔CPU MoE round-trips.
  8 steps × 30 layers × 2 transfers/layer
  = 480 PCIe transfers per canvas.
  Reduce --max-denoising-steps to trade quality for speed.

[arielge]

Haven't gone over the whole adapter logic, but generally - try to keep to a minimum changes to general API unless they are absolutely necessary

[arielge]

model is not gated. Regardless, it should be explicitly excluded from the CAUSAL_PATHS, like we do for dspark entries

[AnishPahilajani]

fyi... 2 card does not work yet. So pls dont merge this.

torchrun --nproc_per_node=2 scripts/test_diffusion_gemma.py --model /models/diffusiongemma-26B-A4B-it --tp --dtype float16 --batch-size 2 --max-denoising-steps 8 --max-new-tokens 256

[BenjSz]

fyi... 2 card does not work yet. So pls dont merge this.

torchrun --nproc_per_node=2 scripts/test_diffusion_gemma.py --model /models/diffusiongemma-26B-A4B-it --tp --dtype float16 --batch-size 2 --max-denoising-steps 8 --max-new-tokens 256

So, for clarity, please move this PR to be draft.