[easy-ish] PPX: replace [%lwt ...] by [%lwt.catch ...] or remove it

[aantron]

The catch-all [%lwt ...] syntax has been a source of problems in #307 and #525, due to how PPX works.

In particular, let%lwt x = e in e' is represented as [%lwt let x = e in e'] in the AST, and all the other special ...%lwt keywords (try%lwt, etc.) are represented similarly, i.e. as ordinary, unadorned expressions nested inside an [%lwt ...] extension point.

Besides keywords, the Lwt PPX provides that if you have [%lwt e], then it will be expanded to Lwt.catch (fun () -> e) (fun exn -> Lwt.fail exn). However, because this conflicts with interpreting the keywords (above), this works for all expressions except let, try, etc., and only for them.

This has two problems:

1. It is a bit difficult to predict and reason about. It's probably a bit scary to refactor, as an exception handler could suddenly disappear if the top level becomes a keyword, though it doesn't have many users.
2. It slows us down adding interactions between %lwt and additional keywords, such as begin%lwt (Ppx: add support for begin%lwt e1; e2; end in ppx extension #525) and ;%lwt (Support ;%lwt as sequence syntax -- backwards incompatible in some cases #307), because giving those keywords a new special interpretation under %lwt can break the interpretation of already-written user code.

This issue proposes to delete the catch-all interpretation of [%lwt ...] from the PPX. If we need a replacement, I suggest [%lwt.catch ...].

I see [%lwt ...] being used in ocsigen and sqlexpr, perhaps as a result of conversion from Camlp4.

Given how few users there are, I suggest to delete this interpretation entirely.

I think we should do this regardless of whether we merge #525 or #307, because it will simplify Lwt, and give it more options for evolving in the future.

cc @Drup @hcarty @kandu @raphael-proust