Tentative fix for #25654

One effect of the simplification in #25105 was that it changed the c-ares resolver to invoke RegisterOnReadableLocked on all active fd's also now when SRV records are resolved. Notably, it does this from a callback that is invoked from the c-ares library during a call to ares_process, and it also clears the readable_registered_ flag before doing this. After this ares_process call completes, we check if there are still more bytes to read on the socket and then continue looping on ares_process if so.

On Windows, RegisterOnReadableLocked pre-allocates space for read_buf_ to read in bytes asynchronously via RecvFrom. Data is actually filled into read_buf_ only when the RecvFrom call completes and the OnIocpReadableLocked callback is invoked. So because we are calling RegisterForOnReadableLocked from a callback invoked from ares_process, when ares_process returns, read_buf_ has a non-zero length but doesn't actually have any readable data. Thus the loop where we continue calling ares_process so long as IsStillReadableLocked returns true will never finish.

read_buf_has_data_ is set true when we read in data from RecvFrom, and is then cleared after c-ares reads all received bytes out of it (this is the proper flag to check if there are still bytes remaining to be read).