feat(moe): add stacked-buffer fast path for SSD streaming (#35)

Introduces `MLX_MOE_STACKED` and `MLX_MOE_FUSE_GATEUP` fast paths to SwitchGLU for SSD-streamed MoE inference. Replaces multiple per-expert kernel dispatches with a single batched gatherQuantizedMM per projection, drastically reducing CPU→GPU enqueue overhead on Apple Silicon.

- Defaults to legacy behavior unless env flags are set
- Automatically and safely falls back if the layer is ineligible (e.g. non-quantized weights, or batch size > 32)
- Added unit tests to ensure fallback safety

Author: solderzzc

.github/workflows/ci.yml

@@ -123,6 +123,12 @@ jobs:
           # Replace the local submodule with the actively checked out code
           rm -rf mlx-swift-lm
           cp -r ../mlx-swift-lm .
+
+      - name: Update mlx-swift submodule to latest main
+        run: |
+          cd SwiftLM/mlx-swift
+          git fetch origin main
+          git checkout origin/main
           
       - name: Build SwiftLM
         run: |

.github/workflows/downstream_integration.yml

@@ -33,6 +33,14 @@ jobs:
           mkdir -p SwiftLM/mlx-swift-lm
           find . -mindepth 1 -maxdepth 1 ! -name "SwiftLM" -exec cp -r {} SwiftLM/mlx-swift-lm/ \;
 
+      - name: Update mlx-swift submodule to latest main
+        # SwiftLM's mlx-swift submodule pin may lag behind main.
+        # Pull latest so new APIs are available for the build.
+        run: |
+          cd SwiftLM/mlx-swift
+          git fetch origin main
+          git checkout origin/main
+
       - name: Install Metal Toolchain
         run: xcodebuild -downloadComponent MetalToolchain || true