Optimizations for Keccak memory machine and adjacent sub machines

[qwang98]

Things I've thought of/heard of so far:

1. One of the following to replace add_sub submachine calls
   a. Implement add4 logic in main machine, putting carry in unused cells of the main machine
   b. Usestd::machines::small_field::pointer_arith::increment_ptr instead

2. Fix many links to the memory machine (a language fix).

3. (Suggested by @georgwiese but I'm not sure how) Fix memory machine, which currently uses one additional fixed column for each permutation call, and another accumulation column overall.

[qwang98]

After implementing 1.b. in #1832 and testing with the same witness generation time as simply using 100 add_sub links, I think 1.a. would have similar results. It's essentially the same as 1.b., but just one fewer carry column (carry will be stored in unused cells in the main machine).

[qwang98]

Update: 1.b. doesn't work in #1832 because we can't disable is_zero constraint in rows other than the first and the last. This is because carry column from increment_ptr implicitly returns a constraint, but isn't explicitly returned by the increment_ptr function, and therefore make_conditional won't disable is_zero constraint properly.

Instead of relying on existing constraint API, I implemented the pointer addition logic within the Keccak main machine itself in #2089.