perf(duplex): lazy projection materialization fixes O(n²) CPU per delta

[Asthenia0412]

Background

Issue #8690 reported that Duplex Realtime projection was rebuilding the cumulative response history on every single audio/text delta, giving O(n²) CPU per response.

Problem

response.output_audio.delta and response.text.delta previously called into _ResponseProjection._rebuild_stored_item() on every delta:

• Joined the full accumulated transcript
• Deep-copied all audio_text_marks
• Deduplicated and sorted the full mark history

For a typical 80-second response with ~2000 deltas, this quadratic behavior dominated CPU.

Solution

Defer the heavy work to the actual consumer read sites. Deltas now only append to the projection's internal accumulators; the stored conversation item is rebuilt only when a consumer actually needs it.

Key architectural changes

• RealtimeProjectionState gains _dirty_response_item_ids (a Python set[str])
• New helpers: _mark_response_item_dirty() (O(1)) and _ensure_response_item_materialized() (lazy rebuild)
• _ResponseProjection gains marks_merged: bool flag and merge_marks() method — dedupe/sort done once at first consumer read instead of per delta
• Delta handlers (response.output_audio.delta, response.text.delta) now only mark the item dirty
• _remember_response_audio_metadata() now only appends raw marks, merge deferred
• Consumer materialization added at:
  • conversation.item.retrieve
  • resolve_truncate_item / conversation.item.truncated / audio.cancelled handler
  • Terminal _response_done_output_item

Performance

Before: N=250 → N=2000 ratio 11.6x   (super-linear, trending quadratic)
After:  N=250 → N=2000 ratio  8.3x   (linear ideal = 8.0x) ✓

Testing

• All existing conversation.item.truncate, .retrieve, and response_done payloads unchanged (pure refactor)
• 16 regression tests pass: dirty flag lifecycle, marks integrity, consumer correctness
• No API surface changes for consumers

Closes #8690