Fix XER collection iteration, data mappings, and round-trip writes

[hurtlovewow]

What changed

• Fix repeated and nested iteration across collection classes (fixes Iterertors outputs works only for one cycle #32).
• Allow the XER reader to accept fields larger than Python's default CSV limit (fixes field larger than field limit (131072) #15).
• Correct TASKPRED project mapping and predecessor lookup behavior.
• Isolate task relationship queries per Reader instance instead of using global state.
• Correct resource, role, category, schedule-option, task, and assignment field mappings and collection lookups.
• Parse ROLE records and preserve role/rate/resource-category tables during XER write/read round trips.
• Harden DCMA14 analysis for empty schedules, missing dates, valid hard constraints, and zero-denominator percentages; remove debug output.
• Add regression coverage for the above, including existing floating-point lag_hr_cnt behavior.

Root causes

The collection iterators retained mutable cursor state between iterations; the reader relied on the CSV module's 131,072-character default field limit; several model constructors and find_by_id() methods contained copy/paste column or identifier errors; and some parsed XER tables were omitted by the writer. DCMA14 also divided by zero and compared dates when one side was missing.

Validation

• pytest -q — 44 passed
• python3 -m compileall -q xerparser tests
• git diff --check

The feature requests for activity-relationship editing (#7), DCMA-14 visualisation (#10), and EVA/CPI/SPI charts (#13) are intentionally not included in this bug-fix PR.