refactor(kernels): the Marlin face serves both consumers from one module

[FeathBow]

Description

Closes #1019

The gemma4 and Kimi Marlin wrappers each carried the same buffer-to-pointer plumbing for the grouped GEMM and the align launch. A shared ops/marlin_face.rs now owns that face, generic over the scale type (u8 for the NVFP4 face, Half for WNA16 — an alias of the same width, so the shared pointers feed both FFIs unchanged); the launch functions consume their buffer structs by destructuring and hand each consumer a pointer set for its own FFI call. The FFI entry points, argument arithmetic and stream selection stay with their consumers, and no exported symbol changes name or signature.

Test Env

Single GPU (sm_89, 48 GiB, x86_64), the kimi-k2 leg compiles with an sm_90 target override, which the moe chain's DeepEP shim requires.

Verification

• cargo fmt --all -- --check and git diff --check pass.
• The gemma4 consumer: cargo clippy --release --features gemma4 -p pegainfer-kernels --all-targets -- -D warnings passes at the exact head.
• The Kimi consumer: PEGAINFER_CUDA_SM=90 cargo check --release --features kimi-k2 -p pegainfer-kernels passes at the exact head, 0 errors.
• Zero symbol renames by construction: the diff touches no pub fn signature, no extern declaration and no entry-point string; consumers outside the two wrapper files compile without edits.