LICM and load CSE: two roots are two objects only when their provenan…

[darmie]

…ce says so

A parameter, a phi, a loaded pointer or a call's result may name the same object as any other such root, in any function, not only in an outlined region. Only allocations the function makes, globals, and objects of two different exact struct types are kept apart.

git-bug: 96fe0b9665e4a9476eaf1ff4421ab1f54b1417ca6323e3ba3a820ee6dcdf7bcc

[darmie]

native tests (x86_64-pc-windows-msvc) fails on fae99be in Python warm_up::exceptions: exceptions/try_finally_tuple_return.py prints the right output, then zypy's stderr ends with "a function compiled on first call could not be compiled (bead 1)" and the process exits 0xC0000409. This is the Windows-only warm-up fault already on main (the same case failed the same way on main's PR #7 runs at 6bb5a3c and 3b26b34, before this PR's change existed), tracked in git-bug b23d6a1; it is not from this PR's LICM change, and every other check here, Linux and macOS native tests included, is green. No fix exists yet; a separate branch is working on it, and once it lands on main this PR only needs main merged in. Not re-running: the failure has already reproduced identically on three Windows runs.