LLVM test-suite failures

As discussed in #25762, and seen in !13970, there are many tests currently failing in the LLVM way.

Both llvm and optllvm ways:
"AtomicFetch CgStaticPointers CgStaticPointersNoFullLazyness GcStaticPointers ListStaticPointers T12622 decodeMyStack decodeMyStack_underflowFrames t24005 tough tough2 staticcallstack001 staticcallstack002 hpc001 hpc_fork"
I analyse these failures in this comment. They reveal that static pointers, IPE information and HPC are not working properly with the LLVM backend.

Linux only: "UnboxedTuples", a serious run-time crash tracked in #25770 (closed).
Darwin only: "jspace" (this is #25401).

llvm way only: "T22187_run" (this is fixed by !13936 (closed)).

optllvm way only: "GenDerivOutput GenDerivOutput1_0 GenDerivOutput1_1 LinearListComprehension T10604_deriving T17574 T18052a T18118 T19381 T20275 T21755 T22744 T24224 T24726 drv-empty-data dsrun014 stack_big_ret"

These failures are unrelated to the LLVM backend, and due to the fact that we run the tests with optimisations. They do reveal genuine bugs, some quite serious:

• T20275 is tracked in #25771 (closed) (GHC crash in constant folding code)
• LinearListComprehension is tracked in #25772 (Core Lint linearity failure).

Linux only: "T14251". Incorrect runtime result with the LLVM backend and optimisation (tracked in #25773 (closed)).