f32<->f16 convert

Eight independent vec2<f32> chains per thread, unrolled 4x: pack2x16float then unpack2x16float (round-trips through fp16 bits) plus a vec2 FMA to keep the chain moving. Counted as 2 ops per lane per step (one convert each way; the FMA is not counted). Unlike f16-*, this needs no shader-f16 device feature — it measures the conversion, not f16 compute.

Measured in OP/s. Higher values mean better performance. Benchmark your device.