Cut 17.0.0 and bump the package version in pyproject.toml, uv.lock, and
CITATION.cff.

Major: the parallelism surface is removed (#142) - the `--threads` / `-t`
flags on `build-kg` and `build-fullmap`, the agent command's
`--gepa-threads` flag, and the `threads` / `num_threads` parameters on
the Python and Rust pyfunctions are gone; migration is to drop those
arguments, since the automatic behavior is exactly the previous default
(all CPU threads, fullmap builds memory-capped on Linux, lookups
parallelized for batches of 1024+ terms) and results are unchanged.
Also shipping: the `uuid_on_collision: merge` fold is near-linear and
the canonical serializer allocation-free (#143) - up to 28.6x on
merge-heavy workloads and 14.6% faster in default `error` mode, with
byte-identical output.

Changelog:
- Versioned the Unreleased section as 17.0.0, renamed its `### Removed`
  subsection to `### Breaking Changes`, gave the threads-removal entry
  its PR link and a migration note, and moved the #143 Performance entry
  out of the already-released 16.6.2 section, where it had been misfiled,
  into 17.0.0.

Docs: docs/cli.md, docs/fullmap.md, docs/api/fullmap.md, docs/agent.md,
and examples/agent/README.md were updated by #142 itself; none needed
here - the release commit touches version metadata and CHANGELOG.md only.

Testing:
- uv run pytest -q -> 1243 passed, 3 skipped (96% coverage)
- uv run ruff check . && uv run ruff format --check . && uv run pyright -> clean / 0 errors
- uv lock --check -> up to date
- uv run mkdocs build --strict -> clean
- cargo fmt --check && cargo clippy --all-targets -- -D warnings -> clean (rust/ touched)