[emme1t]

Model finalizers can run after Python clears module globals, causing the cleanup callbacks to raise TypeError: 'NoneType' object is not callable or AttributeError while releasing native resources.

Capture the native release functions when registering the ExitStack callbacks for models, contexts, batches, samplers (including clones/custom samplers), and LoRA adapters. The callbacks retain their existing release order and idempotent close behavior through module teardown.

Fixes #1998.
Fixes #2002.

Validation on Windows CPU, Python 3.12.13, with the native library built from the pinned llama.cpp revision 0c1e57098 using LLVM-MinGW/Ninja (GGML_NATIVE=OFF, GGML_OPENMP=OFF, _WIN32_WINNT=0x0A00):

• Full python -m pytest tests -q: 32 passed, including the Qwen3.5, Mamba, Falcon-H1, tiny Llama, and embedding-model tests.
• Six new subprocess cases cover explicit/repeated close and garbage collection with intact globals, cleared function bindings, and cleared module references. They verify exactly-once release, context-before-model ordering, and reverse removal of custom samplers. Baseline: 4 failed / 2 passed; fixed: 6 passed. Model/context/batch/sampler resources are real native allocations; the optional LoRA allocation/release is stubbed.
• Real stories260K.gguf inference generated eight tokens before explicit close and garbage collection with the bindings module reference cleared; both completed without unraisable exceptions.
• Repository-wide Ruff lint and formatting checks passed.