feat(cli): create index --train-pq-codebook trains codebooks at creation (#920 sub-item 3)

[mosuka]

Summary

• Second of three sub-items of feat(vector/index): PQ shared-codebook follow-ups (--from-index, FastScan, create convenience flag) (#631 PR-4) #920 (sub-item 1 --from-index merged as PR feat(cli): train pq-codebook --from-index samples committed vectors (#920 sub-item 1) #938; sub-item 2 FastScan remains open). Refs feat(vector/index): PQ shared-codebook follow-ups (--from-index, FastScan, create convenience flag) (#631 PR-4) #920 — does not close it.
• Adds --train-pq-codebook <jsonl> to laurus create index: shared PQ codebooks are trained as part of index creation, so the very first commit can already encode against them. This removes the create → train pq-codebook → ingest ordering hazard that the feat(vector,cli,server): schema-level pq_codebook_path + laurus train pq-codebook (#631 PR-2) #918 failure policy (hard-error at commit, no silent fallback) otherwise leaves the user to manage manually.
• The laurus library crate is untouched — changes are confined to laurus-cli + docs.

Changes

• train.rs: JSONL collection loop extracted into pub(crate) collect_vectors_from_jsonl (behavior-preserving; the 7 existing train tests are the regression gate).
• cli.rs / main.rs: CreateResource::Index gains train_pq_codebook: Option<PathBuf>.
• create.rs::run_index:
  • Resolves the effective schema first, replicating init_index's recovery rule (an existing schema.toml without store/ wins over the argument/wizard schema); all paths now go through create_index_from_schema.
  • Pre-creation validation: a missing JSONL or zero eligible fields bails before anything is created — no half-initialized index.
  • After creation, reopens the index and trains every eligible field (Hnsw + ProductQuantization + pq_codebook_path, sorted by field name since Schema::fields is a HashMap) with output=None, so the persisted schema and trained file agree by construction.
• Docs (EN+JA): laurus-cli/commands.md create-index section (usage, argument row, one-step example); vector_indexing.md shared-codebook section.

Tests

4 new in create.rs (CLI total 17):

• create_with_train_flag_makes_first_commit_succeed — the direct hazard-resolution regression (.pqcb exists, reopened first commit succeeds).
• create_without_train_flag_leaves_first_commit_failing — control pinning the hazard itself: without the flag the same schema's first commit fails with the pq_codebook_path hard-error. Together these form a RED/GREEN pair attributing the fix causally to the flag.
• create_with_train_flag_rejects_schema_without_pq_field — bails and nothing is created.
• create_with_train_flag_trains_every_pq_field — multiple PQ fields all trained.

Verification

• cargo fmt --check / cargo clippy --all-targets -- -D warnings: clean on stable and 1.97.0.
• cargo test -p laurus-cli --bin laurus: 17 passed (+4). cargo test -p laurus --lib: 1273 passed (library crate untouched).
• markdownlint EN+JA: 0 errors; both mdBooks build.