Shortcut generation for map/bind with lazy small_example

[art-w]

This proposal might not be the right bandage (maybe just expanding the documentation of choose would be better.) I ran into some issues with a recursive generator that was failing to terminate. Reading crowbar source code, there's a size fuel mechanism in place to shortcut the random generation by choosing the first available const/primitive when the generated term grows too much -- but my leaves were wrapped in a map [..] and so the shortcut couldn't trigger (for example, see the xmldiff test where knowledge of the optimization edge-case yields a manual unrolling).

Some random comments:

• small_examples was a list of candidates, but then the shortcut would always pick the first one (such that it could have been an option): https://github.com/stedolan/crowbar/blob/0cbe3ea7e990a7d233360e6a74b1cb5e712501ad/src/crowbar.ml#L246 Perhaps the plan was to select a random value from that list? (but the current behavior has the advantage of not consuming more random bits)
• When provided an empty list, choose would previously crash at this point: https://github.com/stedolan/crowbar/blob/0cbe3ea7e990a7d233360e6a74b1cb5e712501ad/src/crowbar.ml#L252 Perhaps Choose [] could be simplified away by the smart constructors -- but this would push the "non-empty generator" constraint to unlazy/bind, it's not clear that this would be better.
• If desired, the PR could be adapted to prefer the left-most available const/primitive as before, and default to the map/bind otherwise (but I think it's easier to predict the default value when it's always the first element of choose)
• Finally, it's not super clear to me that the shortcut doesn't bias the generator toward uninteresting leaves -- hopefully afl-fuzz is able to avoid triggering it too much?

(unrelatedly, the xmldiff test seems to work on my computer -- however the calendar test crashes very fast!)

[lindig]

I tried to understand how the length of lists is controlled; reading the source I came across the size parameter but could not see how it is initialized. A comment in the code of its role and initialization would be useful.