Interpreter: Catch Invalid_argument more locally

[nomeata]

previously, we would catch Invalid_argument and turn it into a Trap
very coarsly. The coarse catch around interpret_block was added in
1e1757a
but now also catches Invalid_argument exceptions that should not be
caught.

Instead, we need to to try … with precisely the points where we expect
an Invalid_argument to stand for something trapping. In this case
Prim.prim.

Prim.prim can't raise trap itself, as Trap is defined in the two
interpreters, and because it doesn't have accees to the exp.at.

So this wraps invocations to Prim.prim, following similar patters for
binop etc.

It also simplifies the type signature of Prim.num_conv_prim, making it
clear that this is just a self-contained call-by-value function.
Otherwise, we’d still catch Invalid_argument exceptions raised by the
continuation that num_conv_prim invokes.

The same should be applied to Prim.prim, but we can’t, because
Array.tabulate is higher order and inovkes a function. But still, this
is an improvement of the status quo.

See https://dfinity.slack.com/archives/CPL67E7MX/p1594979459058400?thread_ts=1594977500.055800&cid=CPL67E7MX

[dfinity-ci]

This PR does not affect the produced WebAssembly code.

[crusso]

I think the correct thing to do here as @grreif has suggested in the past, is to pass success and failure continuations to the prims and do the right thing. Related to #1530

[crusso]

Changed my mind because I wanted to check something...

I was wondering whether we should actually give OtherPrims access to the failure continuation, but for now this is good.