Function subgroup_f_max 

pub fn subgroup_f_max(val: f32) -> f32

Subgroup (wave/warp) floating-point max reduction.

Returns the maximum of val across all invocations in the subgroup.

On SPIR-V: calls spirv_std::arch::subgroup_f_max. On CUDA: uses warp-level __shfl_xor_sync butterfly reduction. On CPU: returns val unchanged (subgroup size = 1).