After landing #3358 the benchmarks exploded indicating problems with workers and deno_core_http_bench.
This PR dramatically fixes thread/syscall count that showed up on benchmarks. Thread count is not back to previous levels but difference went from hundreds/thousands to about ~50.