Setup and hold are influenced by the logic speed, the amount of internal skew between the clock input and the destination logic, and the skew between the signal inputs to be sampled.
Faster logic will have shorter setup/hold time since it has less delay, and therefore skew. But this comes at the expense of power.
Chip designers also factor in how difficult ...
In the 8-variable case, you need a 256-bit LUT. So you distribute that over four 64-bit LUTs, and use a fifth LUT as a 4:1 multiplexer to select the correct output.
So, for 10 variables, you need 16 of your 64-bit LUTs, and a 16:1 multiplexer, which requires 5 more LUTs.
You can also think of it as four copies of your 8-variable circuit (20 LUTs), with one ...
No closed form expression exists for the capacities.
The best you can do is numerical approximation.
The capacity is the mutual information between Y and X, where Y=X+N, X is the constellation point (in your case a PAM), and N is Gaussian.
Then, the mutual information is I(Y;x) = H(Y) - H(Y|X).
The second term is easy (working it you yourself if you want).