lld: pack Relocation to 16 bytes (fork-local)

[julesjacobs]

Stacked on #38 — this PR's diff includes both commits; review only the second commit ("Pack Relocation to 16 bytes").

Builds on the 24-byte layout by additionally narrowing offset to uint32_t and addend to int32_t, reaching 16 bytes — half the original 32. This caps section-relative offsets at 4 GiB and addends at signed 32-bit, so it is a fork-local layout rather than an upstream candidate.

Removes 16 bytes per input relocation from the linker peak footprint: 0.30-0.60 GB per link on 1-2 GB executables with 19-38M relocations. That is 26-30% of the linker anonymous (heap) peak, and 7-8% of whole-process peak on these binaries, whose total is dominated by the file-backed mmap of the output image; the whole-process fraction rises toward the anon figure for smaller outputs. The anon figure is the one that matters for OOMs: under a cgroup memory limit with swap off, anon pages are unreclaimable, while file-backed pages can be written back and dropped under pressure.

Validated as a drop-in against the current production linker (base commit e3c0054): byte-identical output on 54 real links — 52 executables from 80 MB to 2 GB, a shared object, and a linker-script link. Inputs whose section offset exceeds 4 GiB or whose addend falls outside signed 32-bit are rejected with a fatal error rather than silently truncated; the guard never fires on real inputs (observed values sit orders of magnitude inside the bounds) and correctly rejects a synthetic out-of-range addend that the current linker silently truncates.