perf(streaming): skip build_events and use O(1) buffers on get_final_message drain path

[buddywhitman]

Fixes #1649.

Summary

When a caller uses get_final_message() or until_done() without iterating events, every SSE event still pays the full cost of build_events() - constructing TextEvent / InputJsonEvent / etc. objects that are immediately discarded. On responses with p99 ~4 k deltas this is 100–300 ms of pure overhead per call.

Additionally, accumulate_event has two O(n²) shapes:

• content.text += delta — attribute-target assignment; CPython can't apply its in-place string optimization, so each delta reallocates the buffer
• from_json(json_buf, partial_mode=True) — re-parses the growing JSON fragment on every input_json_delta

Changes (src/anthropic/lib/streaming/_messages.py)

1. _emit_events flag — MessageStream and AsyncMessageStream now carry a _emit_events: bool = True flag. until_done() (and transitively get_final_message()) sets it to False before consuming, causing __stream__ to skip the build_events() call and its yields entirely.

2. accumulate_event(emit_events=False) — when draining:

   • Accumulates text deltas into a __text_buf list; joins once at content_block_stop → O(n) instead of O(n²)
   • Skips per-delta from_json for input_json_delta; does one final parse at content_block_stop → O(n) instead of O(n²)

The iteration path (for chunk in stream:) is unaffected — emit_events stays True so TextEvent.snapshot and InputJsonEvent.snapshot remain correct on every yield.

Tests

tests/lib/streaming/ — 17 passed (was 13; 4 new TestDrainPath tests added).
The new tests assert that get_final_message() without prior iteration returns correct text content and parsed tool-input for both sync and async streams.