Specify kernel version when loading CPU gemm_4bit_forward from the Hub

[jiqing-feng]

Description

Loading the CPU gemm_4bit_forward kernel from kernels-community fails with newer
kernels (>= 0.11.1), which now require an explicit kernel API version (or a Hub
revision) to be passed to get_kernel:

A kernel version or revision must be specified. Use `version=<major>` for a stable
kernel API version or `revision=<branch/tag/commit>` for an explicit Hub revision.

Without it, get_kernel("kernels-community/quantization-bitsandbytes") raises and the
CPU 4-bit fused path silently falls back to the slower native implementation.

Fix

Pass the stable kernel API version=1 (an int, matching the repo's v1 build
branch) to get_kernel:

- gemm_4bit_forward_kernel = get_kernel(
-     "kernels-community/quantization-bitsandbytes"
- ).gemm_4bit_forward
+ gemm_4bit_forward_kernel = get_kernel(
+     "kernels-community/quantization-bitsandbytes", version=1
+ ).gemm_4bit_forward

Notes

• version must be an integer (1), not a string; it resolves to the v1 build
  branch published for this kernel.
• The load remains wrapped in try/except, so environments without kernels
  installed continue to fall back to the native CPU kernel.

[matthewdouglas]

This was on my TODO list, thank you! LGTM!

[github-actions]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.