[LoongArch] Ignore debug uses when folding large offsets (#230290)
A `DBG_VALUE` can currently make `foldLargeOffset` reject an otherwise
foldable offset because the pass uses `hasOneUse()`. As a result,
equivalent MIR with and without debug information can produce different
instruction sequences.
Use `hasOneNonDBGUse()` for the profitability check. Since the fold
erases the offset-producing instructions, mark their debug users undef
before erasing them.
The new MIR test covers the case where a large offset has one ordinary
use and one debug use. Without the change, the offset remains
materialized with `LU12I_W` and `ADD_W`; with the change, it is folded
into `PCADDU12I` and `ADDI_W`, and the invalidated `DBG_VALUE` uses
`$noreg`. The checks were generated with
`utils/update_mir_test_checks.py`.
Tests: `check-llvm-codegen-loongarch` (958 passed), `git clang-format
--diff HEAD`, and `git diff --check`.
Assisted-by: OpenAI Codex 5.6 Sol, co-reviewed by Claude and Yongqiang
Tian.