fix: CPU-only Linux build fails to link kv_inplace_write #2108

Problem

src/lib/mlxcel-core/build.rs:143 compiles ../mlx-cpp/turbo/kv_inplace_write.cpp (#1959/#1961) for every backend. Its InplaceSliceWrite::eval_gpu (kv_inplace_write.cpp:43-65) calls mx::copy_gpu_inplace at line 55, which MLX defines only when a GPU backend is built. A Linux build with no cuda/metal/rocm feature therefore fails at link time, e.g. cargo test --release --lib. Found while validating #2093, which needed a linker wrapper with --unresolved-symbols=ignore-in-object-files. Not re-run for this issue (cold MLX build).

No CI job catches it: the ubuntu-latest jobs only lint, every GB10 job passes --features cuda, and ci.yml:1072 is cargo check, which does not link.

Proposed fix

• In build.rs, next to MLXCEL_BRIDGE_METAL_BACKEND (~172) and MLXCEL_BRIDGE_ROCM_BACKEND (~181), define MLXCEL_BRIDGE_GPU_BACKEND when the resolved metal_backend bool (not the cargo feature, see the comment at ~167-170) or CARGO_FEATURE_CUDA or CARGO_FEATURE_ROCM holds.
• In kv_inplace_write.cpp, guard the <mlx/backend/gpu/copy.h> include and the eval_gpu body on that define; without it, eval_gpu throws std::runtime_error like eval_cpu (lines 38-41) does, keeping that explicit error.
• Rejected: dropping the .file() line. cpp/mlx_cxx_ext.cpp:12 includes the header and the FFI symbol must still link; at runtime cache.rs:1043-1047 already skips the call when default_device_is_gpu() is false.
• CI: add a CPU-only cargo test -p mlxcel-core --profile test-fast --lib --no-run step on GB10 with its own persistent target dir (pattern at ci.yml:657-661). If the warm runtime is too high, put the local command in docs/installation.md instead and say so in the PR.

Acceptance Criteria

• With no GPU feature on Linux, cargo build --release and cargo test -p mlxcel-core --profile test-fast --lib --no-run link (they fail on 33c4505).
• --features cuda build and lib tests unchanged; Metal and ROCm builds compile (CI or a maintainer run). CUDA part done in PR fix: gate kv_inplace_write GPU copy on a GPU backend build #2113 (server build, mlxcel-core inplace/cache:: filters); Metal and ROCm still need a maintainer run.
• The CPU link check is in CI, or documented with the reason it is not. Documented in docs/installation.md; the CI step is deferred to fix: WebUI Activity gate collides with host GPU work and stale incremental cache #2111, which owns ci.yml in this run.

Do it in one PR, run the checks once at the end, and push as few times as possible.

Verification

cargo test -p mlxcel-core --profile test-fast --lib --no-run
cargo test -p mlxcel-core --profile test-fast --features cuda --lib -- --test-threads=1