fix: pass backend="cpu" when requesting CPU gemm_4bit_forward kernel

[SIDDARTHAREDDY8]

Fixes #2064.

Summary
bitsandbytes/backends/cpu/ops.py requested the fused CPU 4-bit GEMM kernel from kernels-community/quantization-bitsandbytes via get_kernel(...) without a backend argument. The kernels package infers the backend from the installed torch build, so on any CUDA torch build the inferred backend is CUDA, every CPU-only build variant is rejected, and get_kernel raises at every process start on hosts with AVX512-BF16 (logged as a warning by the best-effort fallback handler). This one-line change passes backend="cpu" explicitly so the CPU build variant is selected regardless of the torch build.

Test notes

• Verified get_kernel accepts a backend kwarg against the kernels library source (huggingface/kernels, load.py): "The backend to load the kernel for. Can only be cpu or the backend that Torch is compiled for."
• Executed the patched module-level loading block in isolation with a stubbed kernels.get_kernel: it now calls get_kernel(repo, version=1, backend="cpu").
• The issue reporter's own repro shows get_kernel(REPO, version=1, backend="cpu") loads the CPU build variant cleanly where the inferred call fails.
• Full E2E not possible here: this box has CPU-only torch (where the inferred backend is already cpu), and the kernels package is not installed. py_compile passes.