Improve generated code for let assert on the Erlang target

[GearsDatapacks]

Closes #4497 and fixes #4145
This PR improves the generated code for let assert in a couple of ways:

• Instead of pattern matching on the value twice, we match once, extract the variables, then return a tuple from the case expression.
• If a pattern always matches (e.g. a plain variable or tuple pattern) we omit the check entirely and just assign like a simple let assignment.
  This simplification fixes the other bug by default, as we no longer ignore variables in the first pattern match.

[giacomocavalieri]

This entire portion of code is being worked on by the decision tree PR, which already gets rid of needless checks using the decision tree the compiler produces; and merging this would likely cause loads of conflicts with it. I'd love to avoid that! 😁

[GearsDatapacks]

Ah I see. I thought that PR was only for the javascript target. My bad!

[giacomocavalieri]

Ah Gears I'm so sorry I totally messed up! You're right, it is only for the js target 🙇‍♂️
This would not overlap all that much with the decision tree branch. If you want to, though you could still use the decision tree on the Erlang target as well to see if the let assert could be simplified, I'm not sure what would be easier but that's what I'm doing on the JS target if you want to have a look at the implementation

Sorry again for the confusion!

[GearsDatapacks]

All good. I think this implementation is going to be fine for now. I imagine Erlang performs some optimisations using a decision tree also. Perhaps once the plumbing is in place for the decision tree we can improve upon it further

[lpil]

Thank you!!!