I disagree that there's a 2^128 chance that they will collide. Trivially, I can show you a hash that will never collide for up to some N, and that is F(P) = P mod 2^128. This will never collide unless P is more than 128 bits long.
My rationale, above, was that SHA constrains the space to 128 bits. Therefore, for different SHA hashes (ones that haven't already collided), the probability that bcrypt will collide might be smaller (or zero, as in my example above).
In reality it doesn't work like that, I know, but in theory you can't be sure that the probabilities of collision will add (or, well, multiply) up.
Clearly it's impossible to be less than zero. So no matter what you do, Defining H(X) to be G(F(X)) will have strictly more collisions than F(X).
The reason I would argue it's greater than zero is that if a function H existed such that H(X) will never collide for X less than 2^128, it would probably have some cryptographic weakness.
Now clearly if I hash a password F(P) and another password F(Q) there is a 1 in 2^128 chance they collide.
Now imagine we do G(F(P)) and G(F(Q)). We first have the chance of 1 in 2^128 that F(P) == F(Q) which implies G(F(P)) == G(F(Q)). However, that is not all!
We now have a new 1 in 2^128 chance that G(F(P)) == G(F(Q)). So we have (about) a 1 in 2^127 chance that the passwords will collide.
But, none of this really matters. Collisions aren't what you worry about with password hashes.