…edge Lever 1)

Enable running the text encoder and the diffusion transformer in separate
contexts so they are never resident at the same time (peak RAM = max instead
of sum) on memory-constrained devices.

- include/stable-diffusion.h: sd_tensor_raw_t / sd_condition_raw_t +
  sd_precompute_condition / sd_free_condition /
  sd_generate_image_with_precomputed_condition.
- FLUX.2 conditioner-skip: when no text_encoders.llm.* tensors are present
  (DiT-only context) skip building the LLMEmbedder instead of failing the
  tensor load; null-guard the conditioner derefs in setup.
- generate_image refactored to generate_image_impl(..., cond_raw, uncond_raw):
  reconstructs the conditioning into its work_ctx and forwards to
  generate_image_internal, which gained precomputed_cond/uncond params + branch
  (skips get_learned_condition when a precomputed cond is supplied).
- sd_precompute_condition / sd_generate_image_with_precomputed_condition impls.

Verified on host (CPU): encoder-only precompute -> free -> DiT-only generate
produces a coherent image; peak ~2.6GB vs ~4.0GB for a combined context.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

SD_USE_VULKAN builds always picked the Vulkan backend whenever a device
was enumerable, leaving hosts no runtime path to the CPU backend. On
devices whose Vulkan driver initializes fine but deadlocks at the first
compute dispatch (PowerVR DXT-48 / Pixel 10, driver v1.602.x) the only
escape is to skip Vulkan entirely. Check the same env vars the llmedge
JNI already sets (and ggml-backend-reg already honors) before touching
each GPU backend.

…ML_DISABLE_VULKAN/GGML_DISABLE_OPENCL backend opt-out

# Conflicts:
#	src/stable-diffusion.cpp

The master merge advanced ggml a8db410 -> 0ce7ad3, whose ggml-vulkan.cpp
requires SPIRV-Headers (<spirv/unified1/spirv.hpp>) that are not wired into
llmedge's Android NDK include path or the Linux CI toolchain, breaking the
Vulkan build with 'spv has not been declared'. a8db410 is the ggml the merged
sd.cpp source was built and device-verified against (S22: standard, sequential,
FLUX) and is fetchable from the leejet/ggml remote. Only ggml is reverted; the
frontend/libwebm/libwebp pointers stay at the merged commits (they do not
affect the native build and their old commits are no longer on their remotes).
Upstream ggml alignment + SPIRV-Headers wiring is deferred to the upstream pass.

Reverts the temporary CI pin to a8db410. llmedge now wires SPIRV-Headers into
both the Android and desktop/CI Vulkan builds (llmedge-spirv-headers.cmake), so
ggml-vulkan.cpp's runtime SPIR-V patching (namespace spv) compiles. This brings
the nested ggml back in line with the upstream leejet/master merge.

Aligns the recorded nested submodule pointers with the checkouts the
Android build and device testing have been running against.