Feat: Optimize operations for Bonsai on mobile devices

[Juste-Leo2]

The primary goal of this PR is to optimize internal operations to improve performance when running Bonsai on mobile devices.

Note : I'm giving myself some time to implement this at a leisurely pace. Don't expect much from this PR just yet; I'll be experimenting with it.

Below is a quick guide on how to perform a standard inference on CUDA.

Required Models (Downloads)
Please download the following weights before testing:

• Bonsai Image Binary 4B (Q1_0 GGUF)
• Flux2 VAE (Safetensors)
• Qwen3 4B UD (Q4_K_XL GGUF)

Build Instructions
To test this branch, clone the repository and build the project with CUDA enabled:

git clone https://github.com/Juste-Leo2/stable-diffusion.cpp
cd stable-diffusion.cpp
git checkout bonsai_dev

# Initialize and update submodules
git submodule init
git submodule update

# Build the project
mkdir build && cd build
cmake .. -DSD_CUDA=ON
cd ..
cmake --build build -j$(nproc)

Testing / Inference
Once built and the models are downloaded (make sure they are in your working directory), run the following command to test standard inference:

./build/bin/sd-cli \
  --cfg-scale 1 \
  --width 512 \
  --height 512 \
  --steps 4 \
  --seed 42 \
  -p "a cat" \
  -o outputs/cat_test.png \
  --diffusion-model bonsai_image_4b-q1_0.gguf \
  --vae flux2-vae.safetensors \
  --llm Qwen3-4B-UD-Q4_K_XL.gguf

[Juste-Leo2]

I've done an initial code cleanup.
Now I'm going to focus on streamlining the backends to keep the ones most useful for our needs and make the code as universal as possible.
So I've decided to keep:

• CPU (universal)
• Vulkan (universal)
• CUDA (NVIDIA GPU)
• Metal (Apple ecosystem)
• HIP (AMD GPU)
• SYCL (Intel GPU)

The other backends are either niche, not useful for our application or too complicated to compile

[Juste-Leo2]

Alright, here are the initial performance improvements on the x86 architecture using an i5-13600K: I went from 46 seconds per step to 36 seconds (512*512). I'm going to switch to the ARM architecture.

[Juste-Leo2]

For those who want to try their hand at inference on mobile devices, I've put together a short guide; you'll need at least 4 GB of RAM. For now, the mobile version hasn't been optimized yet.

[Juste-Leo2]

After spending hours trying to get the code to work on Termux with Vulkan, I found myself unable to make it work. I think I went about it the wrong way, so the best thing to do might be to close this PR and start completely from scratch.

It’s a failure, but hey, it’s no problem :).

[Juste-Leo2]

I'm going to start a new project. It's going to take a while, but I think this is the best approach: https://github.com/Juste-Leo2/bonsai.cpp