Failure with pgbench and --disable-thread-safety in ~v16

Hi all,
(Alvaro in CC.)

While running some tests for 8984480b545d, I have noticed that the TAP
tests of pgbench fail when ~16 is compiled with
--disable-thread-safety:
[16:51:10.467](0.004s) not ok 227 - working \startpipeline with serializable status (got 1 vs expected 0)
[16:51:10.467](0.000s)
[16:51:10.467](0.000s) # Failed test 'working \startpipeline with serializable status (got 1 vs expected 0)'
# at t/001_pgbench_with_server.pl line 845.
[16:51:10.467](0.000s) not ok 228 - working \startpipeline with serializable stdout /(?^:type: .*/001_pgbench_pipeline_serializable)/
[16:51:10.467](0.000s)
[16:51:10.467](0.000s) # Failed test 'working \startpipeline with serializable stdout /(?^:type: .*/001_pgbench_pipeline_serializable)/'
# at t/001_pgbench_with_server.pl line 845.
[16:51:10.467](0.000s) # ''
# doesn't match '(?^:type: .*/001_pgbench_pipeline_serializable)'
[16:51:10.467](0.000s) not ok 229 - working \startpipeline with serializable stdout /(?^:actually processed: (\d+)/\1)/
[16:51:10.467](0.000s)
[16:51:10.467](0.000s) # Failed test 'working \startpipeline with serializable stdout /(?^:actually processed: (\d+)/\1)/'
# at t/001_pgbench_with_server.pl line 845.
[16:51:10.467](0.000s) # ''
# doesn't match '(?^:actually processed: (\d+)/\1)'

This ./configure switch has been removed in 17~, and, while I've not
analyzed the problem in details, I am wondering if this points to an
actual bug with \startpipeline in all branches.

Thanks,
--
Michael