docs: LWJGL correction for BAF + upstream contribution tracking

[bernardladenthin]

Summary

Two independent documentation changes, one commit each.

edc0526 — replace JOCL with LWJGL in BAF's GPU/fat-jar notes. BAF's
OpenCL binding moved from JOCL to LWJGL, which left three places naming jocl
as the source of BAF's off-heap memory and of its all-platform native
bundling stale:

• guides/src/CODE_WRITING_GUIDE-21.md — the FFM "not yet adopted" note now
  points at LWJGL's MemoryStack / MemoryUtil as the current off-heap
  allocators.
• policies/fat-jar-release-assets.md — the "never Central" rationale and
  BAF's row in the per-repo table. BAF still needs no classifier split, but
  for a different reason than before: LWJGL ships one natives-*
  classifier jar per platform and they may all sit on one classpath, with
  LWJGL picking the match at runtime, rather than jocl's single all-platform
  jar. The conclusion is unchanged; only the reasoning behind it was wrong.
• crossrepostatus.md — same correction in the fat-jar summary bullet.

beb3dad — track upstream contributions, so the state survives between
sessions:

• subprocess.h/contributions.md (new upstream folder) — PRs
  #104 /
  #105 /
  #106, filed
  2026-08-01, still awaiting a maintainer to approve the workflow runs.
  Records the container/compiler matrix they were verified against, the
  limitations left in deliberately, and the preprocessor pitfalls worth
  carrying forward — notably that __GLIBC_PREREQ in a flat #if is a
  parse error, not a false condition, on non-glibc platforms.
• llama.cpp/contributions.md — per-patch upstream assessment of the
  patches jllama applies downstream: 0001 filed as issue
  #26416, 0002 kept
  downstream for lack of an in-tree beneficiary, 0009 redirected to
  sheredom because scripts/sync_vendor.py owns the vendored file and would
  have silently reverted a local patch. Also notes llama.cpp's AGENTS.md
  rule that agents must not write PR or issue prose.

Test plan

• Docs-only; this repo has no build. Nothing to run beyond the repo's own
  CI (REUSE / link checks).
• Both new .md files are covered by the existing **/*.md annotation in
  REUSE.toml, so no new REUSE entry is needed.