optimize Hashtbl.replace -> Hashtbl.add

[redianthus]

This was showing up a little bit too much in my measurements (the one in expr.ml but I thought changing the others would do no harm).

EDIT: in the end I went with removing the Hashtbl, using a list is fine because 1. we never need to concatenate two lists, as we simply add one symbol at a time and 2. we can simply use List.sort_uniq at the end, which should be fine as there shouldn't be too many symbols.

[filipeom]

Is it because Hashtbl.add is faster in the happy path?

[redianthus]

Yes, if you look at the implementation in the stdlib you'll see it is doing much less work (or maybe it's not in the happy path but simply becaus ewe have collisions and in this case, add is also better than replace).

[filipeom]

Oh cool thanks!