[Pipeline B] Fix positional PDF field-matching in Filler (#642)

[abhishek-8081]

Closes #642.

Filler.fill_form matched extracted values to PDF widgets by positional index (values in
field-dict order zipped to widgets sorted by physical position). This worked only because
the field-dict order coincidentally matched the widget order — if they ever diverged, values
landed in the wrong boxes on emergency reports.

Now matches by field NAME: for each widget, its name (_pdf_text(annot.T)) is looked up in
the extracted answers. The widget name, the template field name, and the answer-dict key are
the same string (prepare_fillable names widgets after fields; the LLM keys answers by field
name), so name-matching is exact. Unmatched widgets are left unfilled; unmatched answers
aren't drawn. The positional sort is removed — order no longer affects correctness.

New tests (tests/test_filler.py): the key one places widgets in a different physical order
than the field-dict order and asserts each value lands in its correctly-named box — this
fails under the old positional code and passes now. Plus unmatched-widget and
unmatched-answer robustness tests.

155 tests pass (152 + 3 new). ruff clean.

Note: _pdf_text is duplicated into filler.py (with a sync comment) to avoid a circular
import (template → controller → file_manipulator → filler → template) — same pattern as
_resolve_project_file. Worth folding into the path/helper consolidation follow-up.