[WIP][feat](kt-kernel): Add Ascend NPU CPU-MoE backend + llama.cpp b3173 patch support

[Nikonz]

What does this PR do?

Adds an Ascend NPU (CANN) backend to kt-kernel's CPU-expert-offload engine (CPUINFER_USE_ASCEND_NPU), including a CPU-side MXFP4 MoE kernel path, together with the llama.cpp b3173 patches it needs to build (gguf-py NumPy-2 reader fix, GGML_TYPE_MXFP4), auto-applied at CMake configure time.

This is part of a DeepSeek-V4-Flash CPU-offload port to Ascend 910B. Companion sglang-side PR: sgl-project/sglang#TBD (KT wiring for the AscendTP dispatcher).

Box-verified: full kt-kernel build from these exact commits succeeds — Ascend backend selected, both llama.cpp patches auto-applied by the new CMake step against a pristine checkout, .so builds and import kt_kernel_ext succeeds.

Opening as Draft/WIP for early visibility while the companion sglang PR is finalized — not requesting review yet.

Fixes # (issue)

N/A — new hardware backend, not a bug report.

Before submitting

• Did you read the contributor guideline?
• Did you write any new necessary tests? — exercised via a full on-hardware build (Ascend 910B) rather than a new unit test, since this backend requires NPU hardware to build/run; see verification notes above.

[greptile-apps]

Your trial has ended. Reactivate Greptile to resume code reviews.