refactor(core): split mlx_cxx_bridge.cpp by domain; bump mlxcel-core 0.2.0 (#277)

The cxx bridge implementation had grown to 6367 lines in one file. Split the
cleanly-separable specialized blocks into their own translation units, sharing
the few cross-domain helpers (to_shape/to_dtype/from_dtype/gelu_tanh_approx)
via a new mlx_cxx_internal.h. The cxx FFI free functions stay callable across
TUs through the generated header, so only those file-local helpers needed
promotion; no behavior change.

- cpp/mlx_cxx_internal.h: shared MLX/std includes + the dtype/shape helpers.
- cpp/mlx_cxx_kernels.cpp: fused single-token decode Metal kernels (Mamba2/SSM,
  GatedDeltaNet, decode-MoE).
- cpp/mlx_cxx_nemotron.cpp: NemotronH full-forward decode.
- cpp/mlx_cxx_ext.cpp: native safetensors loading + Metal 4 / turbo / paged
  fused-attention launchers.
- cpp/mlx_cxx_bridge.cpp: the remaining general FFI surface, 6367 -> 4616 lines.

build.rs compiles the new files (it already built multiple .cpp via
cxx_build::bridge). mlxcel-core bumped 0.1.0 -> 0.2.0.

Validated: release build clean (fmt + clippy), greedy temp-0 output unchanged
on qwen3-30b-a3b (incl. MLXCEL_FUSED_MOE, the moved kernels) and nemotron-h-30b
(the moved full-forward + SSM/MoE kernels). The remaining bridge.cpp bulk
(ops/quant/norm-rope/attention/activation) is heavily interleaved by domain;
further extraction can follow.

Author: inureyes

Cargo.lock

@@ -1,6 +1,6 @@
 [package]
 name = "mlxcel-core"
-version = "0.1.0"
+version = "0.2.0"
 edition = "2021"
 description = "Low-level MLX building blocks for mlxcel"
 license = "Apache-2.0"

src/lib/mlxcel-core/Cargo.toml

@@ -35,6 +35,13 @@ fn main() {
     let mut bridge = cxx_build::bridge("src/lib.rs");
     bridge
         .file("cpp/mlx_cxx_bridge.cpp")
+        // Bridge implementation split out of mlx_cxx_bridge.cpp by domain
+        // (shared helpers in cpp/mlx_cxx_internal.h): fused decode Metal
+        // kernels, the NemotronH full-forward path, and safetensors loading +
+        // Metal 4 / turbo / paged attention launchers.
+        .file("cpp/mlx_cxx_kernels.cpp")
+        .file("cpp/mlx_cxx_nemotron.cpp")
+        .file("cpp/mlx_cxx_ext.cpp")
         // Fused Sparse-V SDPA kernel launcher. Lives under
         // `src/lib/mlx-cpp/turbo/` so the MLX-upstream-commit upgrade
         // checklist (CLAUDE.md) treats this directory as in-scope.
@@ -125,7 +132,11 @@ fn main() {
     // Rerun if bridge files change
     println!("cargo:rerun-if-changed=src/lib.rs");
     println!("cargo:rerun-if-changed=cpp/mlx_cxx_bridge.h");
+    println!("cargo:rerun-if-changed=cpp/mlx_cxx_internal.h");
     println!("cargo:rerun-if-changed=cpp/mlx_cxx_bridge.cpp");
+    println!("cargo:rerun-if-changed=cpp/mlx_cxx_kernels.cpp");
+    println!("cargo:rerun-if-changed=cpp/mlx_cxx_nemotron.cpp");
+    println!("cargo:rerun-if-changed=cpp/mlx_cxx_ext.cpp");
     println!("cargo:rerun-if-changed=metal/fused_attention_metal4.metal");
     println!("cargo:rerun-if-changed=../mlx-cpp/CMakeLists.txt");
     println!("cargo:rerun-if-changed=../mlx-cpp/patches");