fix(kickstart): decode UTF-8 by hand instead of locale-aware fgetwc()

[mayani]

Summary

Under a C/POSIX locale, fgetwc() treats any byte >= 0x80 as an invalid multibyte
sequence and returns WEOF, indistinguishable from real EOF. yamldump() and
printYAMLStatInfo()'s character-counting loop both looped on fgetwc(), so a
single non-ASCII byte anywhere in captured stdout/stderr silently truncated the
rest of the YAML report — while data_truncated was computed purely from byte
size and never reflected it.

• Added yamlgetutf8() / yamlpututf8() in utils.c to decode/encode UTF-8 by
  hand from raw bytes, independent of locale.
• Undecodable bytes (including UTF-8-encoded surrogates, which are never valid)
  are now skipped instead of stopping the dump, and data_truncated is derived
  from a first pass that also tracks whether any bytes had to be skipped this way.
• yamldump() also now re-indents on YAML 1.1's other line-break characters
  (NEL, LS, PS), which previously could pass through raw and de-indent content
  out of the literal block once locale no longer masked the issue (found during
  adversarial review — a job could otherwise inject sibling YAML keys into its
  own invocation record via a crafted stdout/stderr byte sequence).

Verified with manual repro (valid UTF-8 em-dash case from the issue's repro,
invalid-byte case, CESU-8/WTF-8 surrogate case, and a YAML block-scalar
injection attempt parsed back with PyYAML) plus the full pegasus-kickstart
C test suite.

Skipped review findings

Autonomous review (simplify, security, general-correctness, Codex review,
Codex adversarial-review) surfaced a few additional items that are pre-existing
behavior unrelated to this locale bug, judged out of scope for this fix:

• statinfo.c:513-533 — the two-pass fd counting/skip loop is duplicated and
  now slightly costlier per call than the old fgetwc version; predates this
  change (same double-pass shape existed before), a single-pass rewrite is a
  separate cleanup.
• utils.c:88 yamlgetutf8 — its 3-state return (0/1/2) could be collapsed to
  a bool, but that would lose the "skip invalid, keep reading" signal this fix
  relies on; kept as-is.
• statinfo.c:518,536 — two FILE* opened on the same duped fd via fdopen,
  the first never fclose()'d; pre-existing pattern (existed with fgetwc
  too), not introduced by this diff.
• statinfo.c:493-524 — -B/data_section_size is documented as a byte limit
  but the counting loop counts decoded UTF-8 code points, so multi-byte output
  can exceed the configured limit by up to ~4x. This predates pegasus-kickstart silently truncates YAML report on non-ASCII bytes under C locale #2250 (fgetwc
  counted wide chars identically) — making the limit byte-accurate is a larger
  change than this locale fix warrants; worth a follow-up issue.

e2e note

e2e (pipeline 7116) had a single failure unrelated to this change: Performance Tests: [014-planner-performance-100k, launch-bamboo-test-planner-only] — a
Java planner timing test (100k-job workflow planning took 1760s against an
upper limit of 1600s), which doesn't touch pegasus-kickstart at all. Confirmed
via trace as CI-runner timing flakiness, not a regression from this diff.

Test plan

• make build-c (pegasus-kickstart, pegasus-cluster, pegasus-keg)
• packages/pegasus-kickstart/test/test.sh — 37/37 applicable tests pass
  (the one skipped test requires the Go pegasus-integrity binary, not
  built by make build-c in this environment; unrelated to this change)
• Manual repro of the original issue's em-dash case, an invalid-byte case,
  a CESU-8 surrogate case, and a YAML injection attempt — all verified
  fixed
• e2e (GitLab pipeline 7116) — green except for the unrelated planner
  timing flake noted above

Closes #2250