Template Metal complex scalar lanes

[PhysicistJohn]

Motivation

MLX Metal currently defines complex storage and arithmetic only through the float-specific complex64_t, while MLX CUDA already uses the lane-generic complex_t<T> = cuda::std::complex<T>. This PR independently gives the Metal backend the same extensible type foundation while preserving all current complex64 behavior and kernel entry points.

It complements upstream #3969, which independently templates the Metal C2C FFT scalar lane. Together, the two narrowly scoped changes establish the orthogonal type and FFT foundations needed to evaluate reduced-precision complex support in MLX. That closes a capability gap with accelerator stacks that already provide reduced-precision complex transforms: NVIDIA cuFFT supports FP16 and BF16 complex transforms, and AMD rocFFT supports FP16 complex transforms.

Neither PR introduces a user-facing dtype or commits MLX to promotion, dispatch, or accuracy semantics; those remain separate design decisions.

Summary

• Replace the float-specific Metal complex structure with lane-generic complex_t<T>.
• Preserve every existing kernel entry point through using complex64_t = complex_t<float>.
• Generalize scalar and cross-complex conversions, arithmetic, and Limits for packed half and bfloat lane types.
• Add compile-time layout and reduced-lane arithmetic checks.

Scope boundaries

This patch intentionally does not add public complex32/complex16 dtypes, host promotion rules, Python API, production half/BF16 kernel instantiations, FFT integration, or FP8. Those should be separate proposals after this primitive is accepted.

The current Metal compiler accepts the template implementation but rejects C++20 concept declarations, so mixed scalar operators use a narrow enable_if alias for equivalent compile-time constraints.

Compatibility checks

Check	Result
Release Metal build	Pass
Minimum-target Metal compile/link (macos-metal2.4, macOS 14.0)	Pass
CTest, serial	261/261 pass
Existing complex64 output	Bit-identical for 8 operators across 65,536 deterministic values
Exported Metal symbols	Exact match: 15,692 lines and identical SHA-256
Existing complex64 latency	Add -0.49%, multiply +0.06%, divide -0.10%; no material regression
Metallib size	+13,440 B (+0.01026%); public metadata and function list unchanged

The metallib increase is confined to module and reflection metadata. No new public kernel symbols are emitted.

Reduced-lane proof

Both complex_t<half> and complex_t<bfloat16_t> are exactly 4 bytes. A linked on-device probe produced the expected (-2, 14) result for both types.

LLVM AIR inspection shows native half and bfloat loads, stores, arithmetic, division, and FMA instructions. The reduced paths do not promote their internal arithmetic to float.

A focused on-device streaming complex multiply-add benchmark over 2^24 elements measured:

Lane type	Median kernel time	Relative to float
float	0.8095 ms	1.00x
half	0.3819 ms	2.12x faster
bfloat	0.3826 ms	2.12x faster

This is a raw-type bandwidth-oriented microbenchmark, not an end-to-end MLX dtype performance claim.