Handling rules that construct pullbacks via mutations of copies of their inputs

OK, I admit I can't find a better title for this. But I feel this should be brought up. See previous related issues in #381 #380 #46

@GiggleLiu keeps wanting to make all the return type of pullback to be the same as its input, that's why he was asking how to fix the problem in #381 for all possible cases. However, I don't think this would work with the current thunk mechanism.

I'm thinking there should be a mutability trait to handle this task, and a BangBang version of setindex! (which is setindex!!) so when the rules writer requires the interface, ChainRules can check and error.

Meanwhile, I think part of the problem of using thunk is because current AD compilers like Zygote are not able to do optimization if the compiler is able to do what thunk does, e.g Yota can figure out the gradient accumulation and unused gradient I believe, then it is not necessary to do it lazily.

cc: @oxinabox @sethaxen

Start by reading the related issues #381, #380, and #46, then inspect the current pullback and thunk mechanisms. Consider how a mutability trait and a setindex!! interface would fit the rules-writing contract. Done would require an agreed design for handling rules that mutate copies of their inputs.