Avoid stack underflow in transaction kernel

As a follow-up of #1902, we should investigate if and where in the transaction kernel we drop more felts from the VM stack than we should.

I have one question for @plafer in this context: Is there a way to inspect the `FastProcessor` state from within or outside to get the current buffer size or probably more precisely, where the "bottom" pointer points?
The main use would be to log this value during events (from the outside) or from within the VM to find places where a value was dropped.

Another idea might be to add a (temporary?) flag on the `FastProcessor` API that, if set, would cause the processor to abort the execution if an element was dropped while the current stack size was already <= 16, e.g. during a `(dyn)call`. _If_ I understand correctly, this would allow us to find the exact instructions (assuming an appropriate `SourceManager` is provided) that caused more elements to be dropped than should be allowed.