Bring the KERMT open-model skill bundle up to date with the KERMT repo main
branch (NVIDIA-BioNeMo/KERMT @ 85f255f):

- Adds the released-model download path: scripts/fetch_released_model.py +
  config/released_model.json, and the --pretrained-release / default-to-
  nvidia/NV-KERMT-70M-v2 flow in the finetune / embed / continue-pretrain
  SKILL.md files.
- Refreshes all 8 SKILL.md, all runner/validator scripts, and the unit tests.
- Regenerates the installable plugin SKILL.md copies.
- Updates .skillsource.json (resolved_sha + files list).
- Preserves the toolkit's Apache-2.0 OR CC-BY-4.0 dual-license identifier on
  all synced files (SPDX headers + SKILL.md frontmatter); KERMT ships
  Apache-2.0 only.
- Keeps the toolkit metadata convention: drops KERMT's owner: frontmatter
  field (no personal owner email in the public repo).

.skillsource.json was hand-updated (the autogenerate sync tool is not in the
public repo); a future generator run should reproduce it.

Signed-off-by: Eva Xue <evax@nvidia.com>