`tryEtaExpandRhs` might throw away useful strictness information for join points

I think currently for join points we trim the demand signature to the join arity. But that seems quite wasteful.

I noticed this while looking at #21429 (closed)

Imagine we have something like this:

[join_arity=1]
f_join = \x -> error

We know error has a signature of <1>bot so for \x -> error we would expect a signature like <A><1>bot.

However tryEtaExpandRhs trims the signature to arity one. So what we get is <A>top.

Which is correct! But we lost the knowledge that f_join foo bar will diverge.

Afaik we use eventual divergence for a variety of inline and floating decisions. So it's bound to make some difference. But I'm not sure how important this is in practice.