Fix iter_rows_keyset NULL hang and silent row loss on ties

[varjoranta]

Summary

Replaces the narrow ValueError fix from #2 with proper compound-cursor pagination. iter_rows_keyset now correctly paginates through NULL values in `by` and through tied non-NULL values; both failure modes are gone.

Re #1 (already closed via #2; this PR replaces that narrow fix with the proper one).

What changed

• Cursor columns extended from `[by]` to `[by] + [p for p in pk if p != by]`. PK tail breaks ties and lets the cursor advance through NULL values.
• `ORDER BY by NULLS FIRST` is now explicit so NULL rows iterate first on every backend (Postgres default would be NULLS LAST; SQLite default already NULLS FIRST).
• WHERE cursor has three shapes, Python-side branched on cursor state:
  • first page -- no predicate
  • NULL region -- `(by IS NULL AND <pk_cursor>) OR by IS NOT NULL`
  • past NULL -- `(by, *pk_tail) > (last_by, *last_pk_tail)`
• OR-group fully parenthesized so filters compose correctly.
• Single-column cursor (by IS pk, single PK) keeps the scalar form `by > $k` -- no one-element row-value syntax.
• The ValueError introduced in Fix iter_rows_keyset infinite loop on NULL by column (#1) #2 for NULL-at-page-boundary is removed.

API change

`sql.select_keyset`'s `last_seen` parameter shape changes from scalar to `Sequence` aligned with the compound cursor. This is a breaking change for direct emitter consumers / `db.compose` previewers; `db.iter_rows_keyset`'s public API is unchanged.

Behavior change for nullable `by` columns

NULL rows now iterate first on Postgres (was implicit NULLS LAST). Non-nullable `by` is unaffected. Users who were hitting the ValueError now get correct pagination instead. No previously-working iteration becomes broken.

SQLite requirement

`NULLS FIRST` needs SQLite 3.30+ (2019). All mainstream environments comply (macOS 3.43, Ubuntu 22.04 3.37, Ubuntu 24.04 3.45+). Documented in README.

Test plan

• 7 new integration tests parametrized x 3 backends = 21 runs:
  • all NULL pagination
  • mixed NULL/non-NULL (issue iter_rows_keyset infinite loop when by column contains NULL values #1 repro shape)
  • tied non-NULL values
  • mixed NULL + ties
  • composite PK tie-break
  • filter + cursor (verifies OR-group parenthesization)
• 11 unit tests covering all three cursor states x single/composite PK x filter precedence x placeholder numbering
• `uv run pytest`: 507 passed, 1 skipped (was 484)
• `uv run ruff check`: clean
• simplify pass applied