Function moe_accumulate_encode_offset 

pub fn moe_accumulate_encode_offset(
    encoder: &mut CommandEncoder,
    registry: &mut KernelRegistry,
    device: &DeviceRef,
    accumulator: &MlxBuffer,
    expert_output: &MlxBuffer,
    src_byte_offset: usize,
    routing_weight: f32,
    n_elements: usize,
) -> Result<()>

Like moe_accumulate_encode but reads expert_output from src_byte_offset.

This enables reading from a slice within a larger buffer (e.g. the down _id kernel output which contains top_k rows of hidden data).