[llvm] [NewPM][CodeGen] Let `ModuleToMachineFunctionPassAdaptor` own `MachineModuleInfo` (PR #87226)

[Arthur Eubanks via llvm-commits]

aeubanks wrote:

I meant in the discourse post that we should get rid of `MachineModuleInfo`, and have a separate `Function` analysis that returns a `MachineFunction` (maybe need some level of indirection like a `std::unique_ptr`). So basically the `FunctionAnalysisManager`'s analysis cache would own the `MachineFunction`.

https://github.com/llvm/llvm-project/pull/87226