Fix interpreter references when loading subword values from other stack chunks.

Summary

@mpickering fixed such references in !13957 (closed) when loading full words.

However there is also code loading subwords in the interpreter which doesn't go through the codepath which checks if a slow load is required.

For example

        case bci_PUSH8: {
            W_ off = BCO_GET_LARGE_ARG;
            Sp_subB(1);
            *(StgWord8*)Sp = *(StgWord8*)(Sp_plusB(off+1));
            goto nextInsn;
        }

To fix this the work done in !13957 (closed) needs to be expanded to also cover stack reads which do not use the SpW macro like the one above..