Increase occurrence counts within recursive continuation handlers

[lthls]

This follows a discussion with @Gbury and @Keryan-dev on unboxing for Classic mode, where it's useful to handle occurrences of a variable inside a loop as if there were multiple occurrences.
It's not clear that it's useful for anything except regular variables, so the implementation may be tweaked to reflect that (and maybe avoid re-allocation in some cases).

[lthls]

A bit of context:
In To_cmm, bindings that are used exactly once are aggressively inlined, to take advantage of the peephole optimisations performed in Cmm_helpers. This is of course restricted to cases where the semantics allow it (preserving evaluation order), but typically any expression with no effects, and at most generative effects, benefits from being inlined.
However, these expressions might not be cheap, so by default we don't want to inline them inside loops. Currently this is done by flushing the delayed bindings before a recursive continuation, as we don't know if any of them might be used in the recursive handler, but it prevents us from inlining the uses that are in the body of the Let_cont.

It gets even more tricky with our work on unboxing as we want to inline some primitives even if it duplicates them at runtime (typically, boxed number allocations and closure projections), but the argument to these primitives must not be duplicated, so we need to know as soon as possible whether the primitive is going to be effectively run once (in which case we can inline the arguments, potentially allowing simplifications later), or if we need to bind the argument and the primitive separately.

This PR would allow us to rely on the fact that a variable marked as having one use in the free names is only used once at runtime, so we never need to backtrack our decisions for loops.

[Gbury]

Note: this patch is included in #807

[mshinwell]

This has been merged as part of #905