Remove explicit ContentDeserializer construction in visit functions

[dtolnay]

This is a good suggestion from #2821 (review). I was not able to measure any binary size improvement — the rlib is 111K before and after. But it's worth making the change.

[dtolnay]

I also tried adding:

impl<'a, A, B> Pair for &'a (A, B) {
    type First = &'a A;
    type Second = &'a B;
    fn split(self) -> (&'a A, &'a B) {
        (&self.0, &self.1)
    }
}

to make this function go away, like in the visit_content_map case. See #2823.

But somehow that made the rlib go back up from 111K to 124K, negating the benefit from #2821. I do not fully understand why this happens.

[Mingun]

If you are worried about the size of the generated code, then maybe merge

• Do not generate unique __Field for each struct #2564 (besides, it reduces compile time)
• Always delegate returning errors to the Visitor from Content[Ref]Deserializer #2811 (besides, it fixes one bug and corrects areas of responsibility of Content[Ref]Deserializer). This one removes a lot of generic code, so should expect reducing size, especially when flattening or internally or adjacently tagged enums is used

I rebased both PRs on current master for easing of testing.

[Mingun]

Some tests on my computer (the same test library project as in #2821):

Commit	Size (bytes)	Diff from master (bytes)
master	143694
visitor-driven	152842	+9148
common-field	136930	-6764
both	142462	-1232