fix(glm52): gate kernels behind explicit feature

[xiaguan]

Summary

• add an explicit openinfer-glm52/glm52 feature and make the kernels dependency optional
• forward openinfer-server/glm52 into the model crate feature so GLM5.2 still opts into openinfer-kernels/glm52
• gate GLM5.2 kernel-only Rust modules behind the model feature
• add FlashMLA nested CUTLASS includes for the GLM5.2 TRTLLM FP8 CUDA translation unit, matching the existing glm52 submodule contract

Validation

• cargo tree --workspace -e features -i openinfer-kernels shows no glm52 / moe feature on the default workspace path
• cargo tree -p openinfer-server --features glm52 -e features -i openinfer-kernels shows openinfer-server/glm52 -> openinfer-glm52/glm52 -> openinfer-kernels/glm52 -> moe
• cargo fmt --all --check
• git diff --check
• cargo check --release --workspace
• cargo test --release -p openinfer-glm52 --lib
• OPENINFER_NCCL_ROOT=/data/opt/nccl-2.30.4 cargo check --release -p openinfer-glm52 --features glm52
• OPENINFER_NCCL_ROOT=/data/opt/nccl-2.30.4 cargo check --release -p openinfer-server --features glm52
• OPENINFER_NCCL_ROOT=/data/opt/nccl-2.30.4 cargo test --release -p openinfer-glm52 --features glm52 --lib
• pre-commit hooks passed with OPENINFER_NCCL_ROOT=/data/opt/nccl-2.30.4

Note: cargo test --release --workspace --lib compiled and started the default workspace path with glm52/moe disabled, then failed on local GDR setup in openinfer-comm-cuda-lib::test_gdr::gdr_copy_flag_GPU (Failed to create GDR copy handle).