[lld] [lld][ELF] Fix a corner case of elf::getLoongArchPageDelta (PR #71907)

[via llvm-commits]

heiher wrote:

What if we could get the page offset of current instruction and `pcalau12i` and record in unused immediate bit fields of `addi.d`/`lu32i.d` and `lu52i.d`. I think we can do it correct and these 4 instructions can be scheduled within the 16M(+/-8M) range. This is enough even for a larger basic block.

```asm
.L1:
  pcalau12i $a1, 0        # R_LARCH_PCALA_HI20
  addi.d    $a0, $zero, A # R_LARCH_PCALA_LO12
  lu32i.d   $a0, B        # R_LARCH_PCALA64_LO20
  lu52i.d   $a0, $a0, C   # R_LARCH_PCALA64_HI12
  add.d     $a0, $a0, $a1
```

### A/B/C

page(4k) offset of current instruction and `pcalau12i`.

```
A = ((. >> 12) - (.L1 >> 12))
B = ((. >> 12) - (.L1 >> 12))
C = ((. >> 12) - (.L1 >> 12))
```

### PC

```
pcalau12i: PC = PC
addi.d:    PC = PC - (A << 12)
lu32i.d:   PC = PC - (B << 12)
lu52i.d:   PC = PC - (C << 12)
```

### R_LARCH_*

```
DELTA = DEST - PC[63:12][000]
DELTA = DELTA + (DELTA[11] << 12)
DELTA = DELTA + (DELTA[31] << 32)
DELTA = DELTA - (DELTA[11] << 32)

R_LARCH_PCALA_HI20   = DELTA[31:12]
R_LARCH_PCALA_LO20   = DELTA[11:00]
R_LARCH_PCALA64_LO20 = DELTA[51:32]
R_LARCH_PCALA64_HI12 = DELTA[63:52]
```



https://github.com/llvm/llvm-project/pull/71907