Restrict native shuffle ops to Float32, Int32, and UInt32 for Metal

[WilliBee]

Continuing from #5 (review)

• restricted native shuffle ops to Float32, Int32, and UInt32 for Metal
• left the Float64 Shuffle Up test deactivated as Metal Arrays do not support Float64
• reactivated ComplexType @shfl test for Metal by creating a ComplexTypeMetal with Float16 instead of Float64

[epilliat]

Looks good — small, scoped, and CI-green. Removing the 8/16-bit Metal fast paths is the right call: Apple's AIR shuffle intrinsics are 32-bit only, so the previous code was emitting ccalls to non-existent symbols. The recursive widening in src/warp.jl:122-137 transparently routes narrow types through UInt32, so every bitstype that worked before still works — just one extra convert/reinterpret hop. The new ComplexTypeMetal test (Float16 in place of Float64, since MtlArray doesn't support FP64) is exactly the right coverage for the nested-struct path. Thanks!