In all of these cases, these substitutions are reasonable because they rewrite the recurrence as one of the form S(k) = aS(k - 1) + f(k). These recurrences are often easier to solve than other recurrences because they define S(k) purely in terms of S(k - 1).
Let's do some examples to see how this works. Consider this recurrence:
T(n) = 1 (if n ≤ 1)
T(n) = 2T(n/4) + sqrt(n) (otherwise)
Here, the size of the problem shrinks by a factor of four on each iteration. Therefore, if the input is a perfect power of four, then the input will shrink from size 4k to 4k-1, from 4k-1 to 4k-2, etc. until the recursion bottoms out. If we make this substitution and let S(k) = T(4k), then we get hat
S(0) = 1
S(k) = 2S(k - 1) + 2k
This is now a recurrence relation where S(k) is defined in terms of S(k - 1), which can make the recurrence easier to solve.
Let's look at your original recurrence:
T(n) = a (for n ≤ 2)
T(n) = 8T(n/2) + bn2
Notice that the recursive step divides n by two. If n is a perfect power of two, then the recursive step considers the power of two that comes right before n. Letting S(k) = T(2k) gives
S(k) = a (for k ≤ 1)
S(k) = 8S(k - 1) + b22k
Notice how that S(k) is defined in terms of S(k - 1), which is a much easier recurrence to solve. The choice of powers of two was "natural" here because it made the recursive step talk purely about the previous value of S and not some arbitrarily smaller value of S.
Now, look at the last recurrence:
T(n) = 1 (n ≤ 1)
T(n) = 2T( (n-1)/2 ) + n
We'd like to make some substitution k = f(n) such that T(f(n)) = 2T(f(n) - 1) + n. The question is how to do that.
With some trial and error, we get that setting f(n) = 2n - 1 fits the bill, since
(f(n) - 1) / 2 = ((2n - 1) - 1) / 2 = (2n - 2) / 2 = 2n-1 - 1 = f(n) - 1
Therefore, letting k = 2n - 1 and setting S(k) = T(2n - 1), we get
S(n) = 1 (if n ≤ 1)
S(n) = 2S(n - 1) + 2n - 1
Hope this helps!