Fix mx.remainder floored-mod for float16/bfloat16 on CPU - #3976

mx.remainder returns the wrong result for float16 and bfloat16 on the CPU backend.

The SIMD kernel sign-corrects the remainder so it takes the sign of the divisor, like float32 and the GPU backends. The correction is gated on std::is_signed_v<T>, which is false for the custom float16_t and bfloat16_t, so the half types skip it and return the raw IEEE remainder.

Example with a = [-1, -1, -1], b = [3, 5, 7]:

• float32: [2, 4, 6] (correct)
• float16 and bfloat16: [-1, -1, -1] (wrong)

The correction now gates on !std::is_unsigned_v<T> && !is_complex<T>, in the scalar path (base_simd.h) and the Accelerate SIMD path (accelerate_simd.h). For every type that was already correct this is the same branch, so only the half types change. test_remainder is extended to float16 and bfloat16.

Verified on macOS (M3 Ultra, Metal + Accelerate) and Linux (CPU): float16/bfloat16 remainder now matches float32 on both CPU and GPU, and test_ops.py passes.

complex64 remainder is also wrong on CPU (it drops the imaginary part), which I'll address separately.