[SR-2905] Linux libpwq: Dispatch queues + blocking tasks = poor performance

Issue Description:

Consider this code:

for i in 1...100 {
        DispatchQueue.global().async {
                print(i)
                sleep(1000)
        }
}
dispatchMain()

Environment: Linux, 8 cores

Expected outcome: The first 8 work items start running. When the threads block on the sleep system call, the manager creates additional threads so that the number of active threads is always 8.

Actual outcome: The first 8 work items start running. Then every 1 second, one new work item runs. It takes 93 seconds for all the print statements to run.

There are actually two issues here.

The first issue is, on every run of the manager, the manager only creates 1 additional worker thread even though all of the existing threads are blocked. Instead, it should create min(num_blocked_threads, cpu_count) worker threads so that we always have cpu_count active threads.

The second issue is the manager only runs once a second. I assume this is to prevent the manager thread from taking away significant CPU time from other threads. Is it possible to achieve the same result by using thread priorities? In other words, can we set the manager thread priority to a special number such that the thread scheduler will only schedule it when the CPU is idle?

Run the supplied DispatchQueue.global async/sleep reproduction on Linux and compare its timing and printed output with the stated expected behavior. No file or test is named; inspect the libdispatch worker-thread manager and its scheduling interval. Done means the reproduction no longer exhibits the reported delayed dispatch and the two manager concerns have clear, tested behavior.