[PowerPC][SDAG] `fabs` on known bits of `ppc_f128` is wrong

The program below gives different results for clang with `-O0` versus `-O2`, where the `-O0` result agrees with GCC. So, LLVM optimizes incorrectly on LE targets.

https://godbolt.org/z/sz5n5qKj9

I'm pretty sure this is wrong, the sign bit changes position on BE versus LE, and on LE it is not the most significant bit of the 128-bit value, but only of the hi double (bit 63 on LE, bit 127 on BE).

https://github.com/llvm/llvm-project/blob/83d95d89c520656306130b461252cd0ed9b4f542/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp#L4160-L4164

```c
/// Make this value non-negative.
void makeNonNegative() {
Zero.setSignBit();
}

/// Set the sign bit to 1.
void setSignBit() { setBit(BitWidth - 1); }
```

does not take the endianness into account.

Reproduce the O0/O2 difference from the linked Compiler Explorer example, then inspect llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp around makeNonNegative and setSignBit. Compare the sign-bit handling on little- and big-endian PowerPC targets, and verify that the reproducer agrees across optimization levels once the issue is addressed.

Written by the indexing model from the issue text.