Fix invalidations from Pair(obj) methods

[timholy]

xref sbromberger/LightGraphs.jl#1581

These are mostly julia#15276-related, with some julia#36454 thrown in for good measure.

[timholy]

CC @ChrisRackauckas

[KristofferC]

I'll just never use a do block again :)

[timholy]

do blocks are wonderful, so I'd say use 'em if you want. julia#15276 is a bummer for sure, but we should just write some tooling to detect Core.Box automatically. I'm thinking of using it as an excuse to learn some JET internals (aviatesk/JET.jl#233) but truth be told it would be super-easy to add this to MethodAnalysis (just a couple of lines).

[KristofferC]

do blocks are wonderful,

Okay, so let's look at do blocks. In order to clean up a resource after a scope a do block:

• Causes one extra indentation (per object!).
• Requires a tool to find all the cases where it leads to bad type inference and in those cases it causes and extra level of indentation (the let part).
• Gives the code inside the do part different semantics (return behaves differently).
• Makes the code hard to step through in the debugger
• Requires a closure to be created with some gensymmed named, making it hard to precompile

If this wasn't an stdlib I would immediately get rid of every LibGit2.with and replace it with https://github.com/adambrewster/Defer.jl. No scopes, no indentation, no different semantics. Just run a function after scope exits. Exactly what is required.