perf(spanner): optimize row popping in StreamedResultSet

[olavloite]

Eliminates an O(N^2) bottleneck when consuming query results:

• Previously, StreamedResultSet yielded rows by calling iter_rows.pop(0) in a loop. Because Python lists are contiguous arrays, popping index 0 shifts all remaining elements in memory on every row yielded, causing quadratic overhead for larger result sets.
• Replaces pop(0) with direct iteration (for row in iter_rows: yield row) and detaches the list directly (iter_rows, self._rows = self._rows, []), allowing row iteration to run in linear O(N) time at C speed.

Benchmark Performance Verification

Performance verification of this PR branch (spanner-row-popping) compared against the 7-day nightly baseline of the Spanner Python client running on GCE (n2-standard-2, 2 vCPUs, closed-loop single worker):

Results Summary

Benchmark Scenario	Metric	PR #18316 (spanner-row-popping)	Baseline (main)	Delta	Speedup
Narrow Result Set read-narrow-result-set (200,000 rows, narrow columns)	Mean P50 P90 P99	300.82 ms 374.98 ms 474.97 ms 497.50 ms	2,866.85 ms 730.31 ms 2,866.85 ms 2,866.85 ms	-89.51% -48.65% -83.43% -82.65%	~9.5x faster ~1.9x faster ~6.0x faster ~5.8x faster
Wide Result Set read-large-result-set (100,000 rows, wide columns)	Mean P50 P90 P99	2,223.01 ms 2,147.62 ms 2,336.65 ms 2,482.95 ms	2,565.57 ms 2,603.18 ms 2,725.26 ms 2,565.57 ms	-13.35% -17.50% -14.26% -3.22%	~1.15x faster ~1.21x faster ~1.17x faster ~1.03x faster

Observations

• Narrow Result Sets (~90% latency reduction / ~10x throughput): With 200,000 rows and few columns, row popping in StreamedResultSet was the dominant CPU bottleneck. Removing that overhead yields a massive performance leap.
• Wide Result Sets (~13% - 17.5% latency reduction): While complex column deserialization accounts for more relative CPU time in wide rows, the row popping optimization still delivers a consistent ~13% to ~17.5% speedup across Mean, P50, and P90.

[gemini-code-assist]

Code Review

This pull request optimizes row streaming iteration in both the synchronous and asynchronous implementations of StreamedResultSet by replacing an inefficient O(N^2) list-popping loop with a standard O(N) for-loop. It also introduces comprehensive unit tests covering large batches, stepwise consumption, chunk boundaries, early termination, and mid-stream errors. As there are no review comments, I have no feedback to provide.

[olavloite]

(The system test failures are unrelated)