Let's start with #3: Can you always walk the AST in any way you like after it is built? Yes, you can, and the Dragon book has an extensive discussion about how to do that for a given set of attributes; in particular, how to avoid (or detect) infinite regression.
Moreover, there is really no inefficiency in doing it this way. Or, to put it another way, trying to compute attributes on the fly during the parse does not speed things up much if at all, and can even slow things down. Even if it were slightly inefficient, it generally makes the parser much more complicated than it should be, partly because of the artificial hoops which need to be jumped through to perform the analysis online, and partly because it tends to interlace code which doesn't need to be codependent. Separation of concerns is almost always a good programming model.
With that out of the way, and briefly:
What the Dragon book says (iirc, because I still don't have either of my copies at my fingertips) is that adding markers can destroy the LR property of a grammar. This doesn't happen if the grammar happens to be LL. It's not actually necessary for the whole grammar to be LL, but it's a bit tricky to provide a formal description of what it means for a part of a grammar to be LL, although it's possible. Before LALR parser generators were available, there was quite a lot of research into left corner parsing (LCP), which you could look at if you're interested in the concepts. (I don't have references handy but the phrase should still be usable as a search term.)
If you want to introduce a marker into an LR grammar, you need to ensure that the reduction action for that marker does not conflict with any shift action. For that to be the case, each itemset which contains an item whose • is just before the marker must be free of items which would allow a shift of a symbol which could follow the marker. This is very similar to the LL(1) criterion, because it is effectively saying that the parser needs to be able to predict a single unique production at that point in the parse, something which is not normally required in an LR parser. (Strictly speaking, the parser only needs to be able to predict a single unique production for some lookahead symbols -- as I said, those which might follow the marker -- but that fact doesn't usually help much.)
A common way of creating this sort of conflict is when you have two similar productions with different markers before the same symbol. See the bison manual for some specific examples. (Bison automatically inserts markers for you when you put an action in the middle of the rule, which is why that section is called what it is called.)
LR grammars can recognise a strictly larger set of languages than LL grammars. But here, we're not talking about just recognising whether a sentence is in a language or not. The goal of parsing (outside of some formal language theory discussions) is almost always to recursively break the input into parts (which is the root of the verb "to parse") and understand the relationships between those parts.
All the same, LR parsers are, by and large, more capable at parsing as well. For example, the common language of arithmetic expressions can easily be recognised by an LL parser, and most textbooks insist that students work through the chore of
destroying refactoring a simple operator grammar in order to make it possible to parse with a top-down parser. But the resulting LL grammar does not produce the same AST; it has completely lost any information about operator associativity. In order to use the parse, it is necessary to rewrite the AST to recover that information. The fact that an LL grammar cannot produce a parse tree for left-associative operators seems to me like it should count as a weakness.
In any event, if there is an LL grammar which correctly parses a language then you can indeed insert whatever attributes you like wherever you want to insert them. But that grammar will also be an LR grammar, and you will be able to insert markers freely into that grammar without losing the LR property. Unfortunately, it is possible that adding markers to the $LL(1)$ grammar will produce a grammar which is not $LALR(1)$ (since not all $LR(1)$ grammars are $LALR(1)$). But as far as I know this doesn't show up often in practical grammars.
If you want an academic reference, you can look at Beatty, J. C. (1982). "On the relationship between LL(1) and LR(1) grammars", Journal of the ACM, 29 (4 (Oct)): 1007–1022. (doi:10.1145/322344.322350. I got that from a reference in the Wikipedia entry on LL grammars; the reference link is not pay-walled, at least as of today.)