fix: make parameter loading backend-aware

[happyyzy]

Summary

Fix two backend correctness issues in model loading:

• serialize non-host backend uploads while keeping file I/O and host conversion parallel
• carry parameter usage ops into ModelManager, select a compatible buffer with supports_op, and enable CPU fallback only for unsupported graph nodes

This follows llama.cpp's per-op buffer selection. It adds no model, tensor-name, or shape special cases and does not modify ggml.

Previously loader workers entered the same backend upload path concurrently, while Q4_0 token embeddings were placed in an OpenCL SoA buffer that rejects GET_ROWS(Q4_0).

Image evidence

Tested on Adreno 830 with OpenCL, GGML_OPENCL_SOA_Q, Adreno kernels, Q4_0 DP4A GEMM, and direct VAE convolution. The image runs used ggml from llama.cpp commit 7e1e28cae36d41fe7bbe9dae7c9625de6565c063.

Case	Before	After
Flux.2 Klein Q4_0, 512x512
Z-Image Turbo Q4_0, 512x512

The unpatched default four-thread load aborts in clCreateSubBuffer. The before images use -t 1 only to isolate the independent incompatible-buffer error. Patched runs use the default four loader threads.

Reproduction commands

export GGML_OPENCL_Q4_0_DENSE_DP4A=1

# Flux.2 Klein, before
./bin/sd-cli \
  --diffusion-model "$KLEIN_Q4_0_GGUF" \
  --llm "$QWEN_3_4B_Q4_0_GGUF" \
  --vae "$FLUX2_VAE_SAFETENSORS" \
  -p "a lovely cat" --cfg-scale 1.0 --steps 4 --seed 42 \
  -W 512 -H 512 --vae-conv-direct -t 1 -o klein_before.png -v

# Flux.2 Klein, after
./bin/sd-cli \
  --diffusion-model "$KLEIN_Q4_0_GGUF" \
  --llm "$QWEN_3_4B_Q4_0_GGUF" \
  --vae "$FLUX2_VAE_SAFETENSORS" \
  -p "a lovely cat" --cfg-scale 1.0 --steps 4 --seed 42 \
  -W 512 -H 512 --vae-conv-direct -o klein_after.png -v

# Z-Image Turbo, before
./bin/sd-cli \
  --diffusion-model "$Z_IMAGE_TURBO_Q4_0_GGUF" \
  --llm "$QWEN_3_4B_Q4_0_GGUF" \
  --vae "$Z_IMAGE_AE_SAFETENSORS" \
  -p "a lovely cat" --cfg-scale 1.0 --steps 8 --seed 42 \
  -W 512 -H 512 --vae-conv-direct -t 1 -o zimage_before.png -v

# Z-Image Turbo, after
./bin/sd-cli \
  --diffusion-model "$Z_IMAGE_TURBO_Q4_0_GGUF" \
  --llm "$QWEN_3_4B_Q4_0_GGUF" \
  --vae "$Z_IMAGE_AE_SAFETENSORS" \
  -p "a lovely cat wearing black sunglasses, studio photo" \
  --cfg-scale 1.0 --steps 4 --seed 42 \
  -W 512 -H 512 --vae-conv-direct -o zimage_after.png -v

Checks

git diff --check origin/master..HEAD

cmake -S . -B build-pr -G Ninja \
  -DCMAKE_BUILD_TYPE=Release \
  -DGGML_OPENCL=ON \
  -DGGML_OPENCL_USE_ADRENO_KERNELS=ON \
  -DGGML_OPENCL_EMBED_KERNELS=ON

cmake --build build-pr --target sd-cli -j

The Android/OpenCL sd-cli build also passes against the repository's pinned ggml@eced84c8.

Checklist

• I have read and confirmed this PR follows the contribution guidelines.