Fuzzer: Better fuzzing of globals

[kripken]

With this PR we generate global.gets in globals, which we did not do before.
We do that by replacing makeConst (the only thing we did before, for the
contents of globals) with makeTrivial, and add code to makeTrivial to sometimes
make a global.get. When no suitable global exists, makeGlobalGet will emit a
constant, so there is no danger in trying.

Also raise the number of globals a little.

Also explicitly note the current limitation of requiring all tuple globals to contain
tuple.make and nothing else, including not global.get, and avoid adding such
invalid global.gets in tuple globals in the fuzzer.

[kripken]

Here is an example output from this:

 (global $global$32 (ref null $12) (struct.new $12
  (ref.func $1)
  (f32.const -3402823466385288598117041e14)
  (struct.new $8
   (ref.func $1)
   (global.get $global$29)
   (i32.const 127)
   (global.get $global$19)
   (f32.const 0)
  )
 ))

The global.gets there are nested inside struct.news, which is a realistic pattern seen in J2CL output for example, which we did not emit before.

[tlively]

I would expect us to support arbitrary constant expressions as the children of the tuple.make, though. For example, do things break if we allow a global.get child?

[kripken]

Hmm, that should work, but it would require us to check if we are the toplevel or not, at this location. There isn't a nice way to do that atm. Do you think it's worth it?

[tlively]

I do think it's worth doing at some point, but not necessarily in this PR.

[kripken]

I realized it is just as simple to disallow only tuple.make at the toplevel, and pushed that now. With that we can get tuple globals whose values include global.gets.