Fix race conditions on loop._ready_len

[x42005e1f]

This is a simple PR that fixes #720 by removing loop._ready_len in favor of len(loop._ready). The attribute was probably added as a premature optimization (it does not affect performance even in edge cases), but it leads to race conditions both between worker threads calling loop.call_soon_threadsafe() and between any such worker thread and the event loop, since loop._ready_len is updated in a non-thread-safe manner (free-threading).

[fantix]

Perfecto!

[x42005e1f]

The test failure looks like a load on CI/CD. Compared to #719, some jobs run faster, some run at the same speed, and some run slower. The problem with test_process_streams_redirect() is that it also captures asyncio messages in stderr that a task's callback is taking too long (probably due to slow context switching between threads/processes under high load on the system scheduler?), so it is flaky. The warning is logged from _on_idle(), so it is unrelated to the changes in this PR.