fix: type inference regression due to From<T> for Signal<T> impl (closes #3200)

[gbj]

No description provided.

[zakstucke]

As outlined in #3200 (comment), I don't think this is the right way to go.

SimpleCounter<T>(#[prop(into)] step: Signal<T>) will always be a broken pattern (as it was in 0.6 anyway inference-wise), the compiler will always have an impl conflict on some input type that should be conceptually valid.

[gbj]

Have you tried SimpleCounter<T>(#[prop(into)] step: Signal<T>)? It works on 0.6, and it works on 0.7 with this PR (see here) so I'm not sure what you mean by saying it "will always be a broken pattern," unless I'm misunderstanding something.

[zakstucke]

will always be a broken pattern

In that it will always cause problems, with this PR you fix the example for e.g. RwSignal/Signal inputs specifically for the rare generic T in a signal prop interface, but has these 2 downsides:

1. raw e.g. step=5 still fails, requiring step=Signal::stored(5), this was also the case in 0.6
2. most importantly, any usages of #[prop(into)]: Signal<usize> now require a manual Signal::stored(..) for raw usize where they would otherwise not need it, which is really painful downstream as you can imagine this is predominant usage pattern.

With the solution I offer in #3200, 1. is still required and unavoidable, but 2 does not apply anymore and is solved.

(I keep mentioning rare by the way, I back that up by only having used this once myself, and looking upstream at thaw who opened the issue, they only use this pattern once in the entire lib that I could find)

[gbj]

I think you're right, see my comment back on the issue :-)

[zakstucke]

Awesome, glad we hashed this one out, this PR made me nervous 😅