Implement lowering of torch.aten.exponential to linalg.

[godot73]

torch.aten.exponential is in the op-to-implement list.

[godot73]

High-level direction:

When x is sampled from uniform(0, 1), f(x) := - ln(1-x)/lambda produces exponentially distributed random variables. Thus, we can lower torch.aten.exponential(x, lambda) by decomposing it into:

%one_minus_x = sub(%one - x)
%log_one_minus_x = log(%one_minus_x)
%neg_log_one_minus_x = neg(%log_one_minus_x)
%result = sub(%neg_log_one_minus_x, %lambda)