Expose Cmlx product + enable jaccl distributed backend

[anupsv]

Summary

Enables the Swift package layer of MLX to compile the mlx-c/mlx/c/distributed*.cpp C wrappers and the jaccl backend, and exposes the existing Cmlx target as a public library product. This lets downstream Swift consumers (e.g. d-inference provider-swift) link against the C distributed-group API and use jaccl for Mac-to-Mac cluster pipeline inference over Thunderbolt 5.

Upstream context (worth reading before merging)

This change deviates from upstream's stance — the original Package.swift comment reads // do not build distributed support (yet). The "(yet)" tracks ml-explore/mlx-swift#371 ("Add distributed communication framework for multi-device tensor parallelism", open since 2026-03-15), which un-excludes the same files and adds polished Swift bindings (DistributedGroup, MLXDistributed.allSum / .send / .recv / etc., plus sharded NN layers).

We're enabling the backend on our fork ahead of ml-explore#371 merging because the d-inference cluster work already depends on it (commit d1266de3 in d-inference, the encrypted pipeline inference stack). If/when ml-explore#371 lands upstream, we'll rebase our fork onto the upstream API and drop any local Swift bindings that overlap.

Known jaccl issues we accept

Issue	Impact for our use
ml-explore/mlx#3149 — JACCL point-to-point send/recv with varying shape produces wrong data or hangs	Low risk for our shape. d-inference uses jaccl only for small collective-op synchronization; activation/token transfer goes over plain TCP with AES-256-GCM, not jaccl.
ml-explore/mlx#3467 — RTR failure on Apple Thunderbolt RDMA, GID-selection regression	Track upstream fix; affects connection setup.
ml-explore/mlx#3442 — backend="any" picks ring instead of jaccl	We invoke jaccl explicitly, not via "any".

What's in this PR

File	Change
Package.swift	Un-exclude mlx-c/mlx/c/distributed.cpp and distributed_group.cpp (the C API wrappers that back mlx_c_distributed_group_*); un-exclude jaccl backend sources (jaccl.cpp, mesh.cpp, ring.cpp, utils.cpp) and exclude the no_jaccl.cpp stub instead; add .library(name: "Cmlx", targets: ["Cmlx"]) to the products: list. Other backends (mpi, ring, nccl) remain excluded.
Source/Cmlx/include-framework/Cmlx.h	Surface mlx-c-distributed_group.h and mlx-c-distributed.h in the Cmlx umbrella header so the distributed-group symbols are reachable when downstream code imports Cmlx.

Net: 2 files, +7 / -8 lines. No behavior change for existing MLX, MLXNN, MLXRandom, MLXFast, MLXOptimizers, MLXFFT, MLXLinalg consumers. Build size grows by the jaccl sources (~600 LOC of C++).

Test plan

• swift build succeeds on macOS 14+ (Apple Silicon)
• Existing MLX consumers compile against this package (verified via d-inference provider-swift)
• Downstream import of Cmlx + mlx_c_distributed_group_* calls succeed in d-inference's ClusterDiscovery / MLXDistributed modules
• Two-Mac Thunderbolt 5 smoke test runs jaccl collective op without hitting #3467