Adjust the FA futility bound with same value as the efficacy bound for asymmetric beta-spending non-binding/binding designs

[LittleBeannie]

The observed gsDesign practice is:

• Calculate beta-spending futility bounds at interim analyses.
• At the final analysis, set the futility bound equal to the efficacy bound, leaving no continuation region.
• After making the bounds equal, it recomputes both crossing probabilities through gsprob(), as shown at line 1612. The probability-complement logic in gsDesign2 achieves the same final partition, but it is not a line-for-line copy of gsDesign’s probability implementation.

Both asymmetric binding and non-binding sample-size paths call gsI1():

• Binding test.type = 3: gsDType3ss()
• Non-binding test.type = 4: gsDType4ss()
  The exact final-bound assignment is in gsI1():

x <- gsBound1(-theta, xr$root * tx, -b, probhi = beta, tol = tol, r = r)
error <- x$error
x$b[k] <- -b[k]
x <- gsprob(theta, xr$root * tx, -x$b, b, r = r)

Here, b is the efficacy boundary. Because the futility boundary passed to gsprob() is -x$b, assigning

x$b[k] <- -b[k]

produces

-x$b[k] == b[k]

at the final analysis.

In this PR, my changes below are the gsDesign2 probability-complement implementation, not a call equivalent to gsprob():

# Assign the remaining probability mass to futility when no final continuation region remains
    if (k == n_analysis && close_final_bound) {
      previous_lower_prob <- if (k > 1) sum(lower_prob[seq_len(k - 1)]) else 0
      lower_prob[k] <- 1 - sum(upper_prob) - previous_lower_prob
    }

It sets the final incremental futility probability so that
$$\sum_{i=1}^{K} P(\text{upper at } i)+\sum_{i=1}^{K} P(\text{lower at } i)= 1,$$
where $P(\text{upper at }i)$ is the unconditional stagewise probability of first crossing the efficacy boundary at analysis $i$, under the value of theta used for the power calculation; and $P(\text{lower at } i)$ is the unconditional probability that the trial first crosses the lower (futility) boundary at analysis $i$.

Therefore, they do not recompute both crossing probabilities. They retain the already calculated upper and interim lower probabilities, then calculate the final lower probability as the remaining probability mass.