perf: Skip fully-resolved sections in subsequent relaxation iterations

[lapla-cogito]

Have collect_relaxation_deltas() distinguish between permanent failures (e.g. unresolvable symbol, invalid register) and
address-dependent failures (e.g. distance too far for jal, value outside c.lui range). Only sections with address-dependent failures are rescanned on subsequent iterations.

[davidlattimore]

Wall time improvement looks to be less than 1%, but some of the other metrics improved.

Benchmark 1 (577 runs): /home/david/save/wild-riscv/run-with env-rand /home/d/wild-builds/2026-02-21.cg1.b --no-fork
  measurement          mean ± σ            min … max           outliers         delta
  wall_time          51.9ms ± 1.07ms    48.7ms … 55.5ms          8 ( 1%)        0%
  peak_rss            173MB ±  779KB     171MB …  175MB          5 ( 1%)        0%
  cpu_cycles         1.24G  ± 23.3M     1.16G  … 1.31G          12 ( 2%)        0%
  instructions       1.55G  ± 14.6M     1.51G  … 1.62G          14 ( 2%)        0%
  cache_references   34.6M  ±  489K     33.1M  … 36.5M          19 ( 3%)        0%
  cache_misses       6.45M  ±  102K     6.15M  … 6.81M          14 ( 2%)        0%
  branch_misses      2.83M  ± 28.9K     2.73M  … 2.95M          12 ( 2%)        0%
Benchmark 2 (581 runs): /home/david/save/wild-riscv/run-with env-rand target/cg1/wild --no-fork
  measurement          mean ± σ            min … max           outliers         delta
  wall_time          51.6ms ± 1.09ms    48.6ms … 56.7ms          8 ( 1%)          -  0.6% ±  0.2%
  peak_rss            173MB ±  779KB     171MB …  176MB          6 ( 1%)          +  0.0% ±  0.1%
  cpu_cycles         1.20G  ± 22.8M     1.14G  … 1.34G          12 ( 2%)        ⚡-  2.6% ±  0.2%
  instructions       1.53G  ± 14.2M     1.49G  … 1.63G          14 ( 2%)        ⚡-  1.4% ±  0.1%
  cache_references   33.9M  ±  542K     31.9M  … 37.2M           7 ( 1%)        ⚡-  2.2% ±  0.2%
  cache_misses       6.15M  ± 97.1K     5.82M  … 6.53M           8 ( 1%)        ⚡-  4.7% ±  0.2%
  branch_misses      2.72M  ± 28.4K     2.64M  … 2.93M          14 ( 2%)        ⚡-  3.8% ±  0.1%

[lapla-cogito]

I'm considering merging the section address calculation logic in finalise_layout() and compute_section_and_symbol_addresses() since a significant portion of their functionality overlaps. This should have a greater impact on wall time than this change.