fix: preserve text after <w:br/> by making CTR run-inner-content attributes collections

[ronaldtse]

Summary

Fixes the plurimath "Bug B" report: text after <w:br/> is dropped on parse.

Per reference-docs/shared-math.xsd lines 421-438, CT_R's content model is:

<xsd:choice minOccurs="0" maxOccurs="unbounded">
  <xsd:group ref="w:EG_RunInnerContent" />
  <xsd:element name="t" type="CT_Text" minOccurs="0" />
</xsd:choice>

The choice has maxOccurs="unbounded", so every element it contains can repeat, interleaved with the others. A math run like:

<m:r><m:t>x</m:t><w:br/><m:t>y</m:t></m:r>

is schema-valid and Word-emitted: text on both sides of a break.

The bug

CTR#t was declared as a scalar attribute (attribute :t, :ct_text, no collection:), so parsing the above XML stored only one t value — the second overwrote the first. Same bug applied to every non-collection attribute in EGRunInnerContent and EGWordRunInnerContent: br, del_text, instr_text, del_instr_text, no_break_hyphen, object, pict, fld_char, ruby, footnote_reference, endnote_reference, comment_reference, drawing.

The previous PR (#8) made br parseable (via CTWordprocessingEmpty/CTWordprocessingText parallel types) but didn't catch that all these attributes needed to be collections for the unbounded choice to work correctly.

The fix

• EGRunInnerContent: attribute :t, :ct_text → attribute :t, :ct_text, collection: true
• EGWordRunInnerContent: every attribute gets collection: true (replaces the inconsistent mix of scalar and collection: 0..1)
• CTR: add ordered declaration to the xml block so the serializer emits children in parsed document order, not grouped by source group. Without ordered, <m:t>x</m:t><w:br/><m:t>y</m:t> would serialize as <w:br/><m:t>x</m:t><m:t>y</m:t> — wrong order, breaks round-trip fidelity and changes document semantics.

Test plan

• bundle exec rspec — 335 examples, 0 failures (was 333; +2 new interleaving tests)
• bundle exec rubocop — clean
• Manual verification: <m:r><m:t>x</m:t><w:br/><m:t>y</m:t></m:r> now parses with r.t.map(&:content) == ["x", "y"] and r.br.size == 1
• Round-trip preserves element order: serialize produces <m:t>x</m:t><w:br/><m:t>y</m:t> (not grouped)
• Updated existing specs that assumed scalar access (.t.content → .t.first.content)

New specs

spec/omml/wordprocessing_in_math_spec.rb adds a "multiple occurrences within one run" describe block:

• "preserves text appearing both before and after <w:br/>"
• "preserves multiple <w:br/> interleaved with text"
• "round-trips a complex run with multiple wordprocessing children"

Impact for consumers

Breaking change for direct attribute access: r.t now returns an array, not a single CTText. Consumers (Uniword, Plurimath) accessing r.t.content need to update to r.t.first.content or r.t.map(&:content).

This breaking change is necessary for schema correctness. The previous scalar form silently lost data on every multi-element math run — any document with text on both sides of a break, multiple breaks, or interleaved wordprocessing elements was corrupted on parse.