Add prefix scan op for Mamba SSM prefill

[esmalTT]

We require a prefix scan operator to implement Mamba SSM prefill. The operation should implement the following torch code:

def sequential_prefix_scan(a, bx):
    (_, _, L, EN) = bx.shape
    hidden_states = torch.zeros((1, 1, L, EN), device=a.device)
    for i in range(L):
        hidden_states[:, :, i] = a[:, :, i] * hidden_states[:, :, i - 1] + bx[:, :, i]
    return hidden_states

Where L is the sequence length and EN = 5120 * 32.