fix(gfql/cypher): preserve openCypher Duration components through toString + equality (#1361) - #1364

Summary

_normalize_duration_map (graphistry/compute/gfql/temporal_text.py) now keeps the three openCypher CIP Duration component groups (months, days, seconds+nanoseconds) separate through canonicalization instead of collapsing days into total nanoseconds.

Behavior changes (canonical strings)

• duration({years: 12, months: 5, days: -14, hours: 16}) → P12Y5M-14DT16H (was P12Y5M-13DT-8H)
• duration({days: 1, milliseconds: -1}) → P1DT-0.001S (was PT23H59M59.999S)
• duration({...}) = duration({...}) now returns false for two durations with equal totals but different (days, seconds) component shapes (was incorrectly true because both sides normalized to the same total-ns value).

Why

openCypher CIP Duration: months, days, seconds+nanoseconds are kept as separate components — months have variable length (28-31 days), and days have variable length under DST. Within a duration, fractional larger units cascade into the next group only (fractional months → days × 30.436875, fractional days → seconds × 86400); whole-unit values do not cascade.

Pygraphistry was previously merging everything into total_nanoseconds and decomposing back via _format_signed_day_time_duration, which lost the days-vs-subday distinction and produced incorrect canonicals for negatives crossing component boundaries.

Scope

Closes 3 of 14 wrong-row cases under #1361 / #1353 item #2:

• expr-temporal6-6-2 — toString({years: 12, months: 5, days: -14, hours: 16})
• expr-temporal6-6-8 — toString({days: 1, milliseconds: -1})
• expr-temporal7-6-8 — equality of two durations with same total seconds but different days+seconds shape

The remaining 11 are tck-gfql repo issues, filed and deferred:

• 10 cases: tck-gfql#38 — port <gt> placeholder substitution gap in Temporal7.py
• 1 case (expr-temporal2-6-5): niche historical TZ for Stockholm 1818 LMT — out of scope.

Test plan

• All 25 existing duration tests in test_lowering.py pass.
• 3 new behavioral tests added at test_lowering.py:4541+ covering the three Bucket C scenarios.
• Full graphistry/tests/compute/gfql/ = 1900 passed, 117 skipped, 15 xfailed — no regressions.
• ./bin/ruff.sh graphistry/compute/gfql/temporal_text.py graphistry/tests/compute/gfql/cypher/test_lowering.py clean.
• CI green
• cuDF dual-version validation on dgx-spark — duration canonicalization is pure-python text formatting (no engine-polymorphic dataframe ops); evaluating whether validation is required or can be skipped.

Cross-reference

• Concrete tracker: GFQL/Cypher: drive down expr-temporal wrong-row tranche (14 cases, #1353 item #2) #1361
• Meta: Meta: Prioritized direct-Cypher wrong-row bugfix tranches (92 cases) #1353 item Don't columns to rename them, use bindings to refer to them instead #2
• Sibling tck-gfql: tck-gfql#36 (in flight) + tck-gfql#38 (filed)