rand: Ensure rd_jitter functions properly for new threads

[brimonk]

In cases where a Windows producer spins up a new thread to produce a single message, the value emitted by rd_jitter is exactly the same for every new thread.

On Windows, calls to srand() modify thread-local state, with subsequent calls to rand() reading that same thread-local state.

The previous fix to call srand() once during rd_kafka_new (#2795) does not fully fix this situation, as a brand new thread can take the object without ever calling rd_kafka_new() or rd_kafka_global_srand().

This patch rectifies this problem by using a version of rand_r() from the musl c library for Win32 platforms. On POSIX platforms, rd_rand_r() is a #define that points to the system's existing rand_r().

This PR addresses the issue I raised in #3929.

[brimonk]

@edenhill the commit history on this PR is a little messy. If you'd like me to clean it up as per the contributing guidelines, I can.

[cla-assistant[bot]]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}