Use non-poisoning locks engine-wide

[james7132]

Objective

Across the engine, we're regularly calling mutex.lock().unwrap(), or mutex.lock().unwarp_or_else(PoisonError::into_inner). It's clear that Bevy, as an project cares little about handling the poisoning. It'd be great if that state could be cleaned up.

Partner PR to #20630. Partially addresses #12660.

Solution

Wrap std::sync locks to expose an infallible interface for both Mutex and RwLock. Unlike parking_lot, these wrappers do not panic upon trying to lock a poisoned lock, and it instead just swallows the PoisonError and silence the poison. This mirrors the current heavy use of uwnrap_or_else(PoisonError::into_error). This is a strong assertion that either all code interacting with the Mutex/RwLock either never panics while a guard is held, or that any panic that may occur leaves the data held by the Mutex/RwLock to be never be left in an inconsistent state where logical (not safety) invariants remain unbroken. This doesn't seem to be an issue with any of our current uses.

As a separate cleanup...

• Both TypeRegistryArc and FunctionRegistryArc have been removed, as users of TypeRegistry and FunctionRegistry can implement their own without too much work
• Removed the explicit functions wrapping DrawFunctions as, opting for a Deref derive instead.
• The internal Arc of AppTypeRegistry and AppFunctionRegistry are no longer pub and no longer implements DerefMut as it just wraps an Arc and it makes no sense to implement.
• SceneLoader just wraps a clone of AppTypeRegistry instead.
• Use SyncCell instead of Mutex where reasonable.

This does result in bevy_platform::sync deviating from std::sync, so alternatively, we can jump the gun and provide std::sys::nopoison before that becomes stabilized and expose these wrapper separately from the poisoning locks.

Testing

Local CI run.