Function adam_step 

pub fn adam_step<T: Triton, const BLOCK_SIZE: i32>(
    params_ptr: T::Pointer<f32>,
    grad_ptr: T::Pointer<f32>,
    exp_avg_ptr: T::Pointer<f32>,
    exp_avg_sq_ptr: T::Pointer<f32>,
    n_elements: i32,
    step_size: f32,
    bias_corr2_sqrt: f32,
    beta1: f32,
    beta2: f32,
    eps: f32,
    weight_decay: f32,
)
where
    T::I32Tensor: Tensor<i32, 1> + Comparison<i32, BoolTensor = T::BoolTensor>,
    T::Pointer<f32>: AddOffsets<i32, 1, T::I32Tensor, Output = T::Tensor<T::Pointer<f32>>>,

Adam step.

exp_avg    = beta1 * exp_avg    + (1 - beta1) * g
exp_avg_sq = beta2 * exp_avg_sq + (1 - beta2) * g²
denom      = sqrt(exp_avg_sq) / bias_corr2_sqrt + eps
p          = p - step_size * exp_avg / denom

Scalars precomputed on host:

• step_size = lr / bias_correction1 where bias_correction1 = 1 - beta1^t
• bias_corr2_sqrt = sqrt(1 - beta2^t)

Grid: [ceil(n_elements / BLOCK_SIZE), 1, 1].