Improve MDL to be scalable with many thousands OS threads

In MDEV-26873, a high contention workload is run with 10K concurrent users using default --thread-handling=one-thread-per-connection. The throughput is very low, and the main bottleneck is find_deadlock() by the MDL.

it does much better with pool-of-threads, even with thread-pool-max-threads=1000, but one can argue that fixing/relaxing this bottleneck could be important.