perf(parquet/pqarrow): avoid goroutines for serial struct reads

[fallintoplace]

Summary

• Use direct child loops in structReader.SeekToRow and LoadBatch when Parallel=false.
• Keep the existing errgroup path for parallel reads.
• Continue visiting every child after an error and return the first error, matching the old serial behavior.
• Add a regression test and a nested struct benchmark.

Benchmark

Apple M1 Pro. upstream/main at 6b039a76. 500ms per sample, 5 samples.

The benchmark reads a 1024-row nested struct across 16 uncompressed row groups and includes SeekToRow(0) plus NextBatch(1024).

children	main	this PR	change	allocs/op
1	120.5 us	103.6 us	14.1% faster	723 -> 715
8	869.1 us	784.0 us	9.8% faster	5708 -> 5672
32	2.927 ms	2.614 ms	10.7% faster	23300 -> 23166
128	10.017 ms	9.140 ms	8.8% faster	93066 -> 92549

Command:

go test ./parquet/pqarrow -run '^$' -bench '^BenchmarkReadNestedStructSerial$' -benchmem -benchtime=500ms -count=5

Checks

• PARQUET_TEST_DATA=<parquet-testing-data> go test ./parquet/...
• go test -race ./parquet/pqarrow -run '^(TestStructReaderSerialOperationsVisitEveryChild|TestRecordReaderSerial|TestRecordReaderParallel|TestRecordReaderSeekToRow|TestPartialStructColumnRead)$' -count=1
• go vet ./parquet/pqarrow
• git diff --check