fix(driver): memory leaks on drop

[johnnyshields]

I used asan to look for leaks in compio-driver and found 3 sources of memory leaks in compio-driver when the driver or its resources are dropped without all operations completing. Found via AddressSanitizer + LeakSanitizer (-Zsanitizer=address with -Zbuild-std).

Before: 12,952 bytes leaked in 18 allocations across the test suite
After: 0 bytes leaked

Changes

1. io_uring Driver — in-flight key tracking + Drop impl

push_raw_with_key calls key.into_raw() to leak the ThinCell<RawOp> allocation as a u64 user_data into the io_uring submission queue. On completion, ErasedKey::from_raw() reconstructs ownership. But if the Driver drops before all completions arrive (which happens on every normal process exit), the raw pointers are never reconstructed and the allocations leak.

Fix: Track in-flight raw pointers in a HashSet<usize>. Remove on completion. In Drop, drain the completion queue, then free any remaining tracked keys via ErasedKey::from_raw().

I've also committed the asan script I used.

Edit: These have been reverted

2. FrozenKey — missing Drop impl

~~FrozenKey wraps ManuallyDrop<ErasedKey> and has no Drop impl. It's used in push_blocking to send operations to the thread pool. If the thread pool drops the closure without running it (e.g., on shutdown), the inner ErasedKey leaks.

Fix: Add impl Drop for FrozenKey that drops the inner ErasedKey. Modify into_inner() to wrap self in ManuallyDrop first, preventing double-drop on the normal (non-leak) path.~~

~~ 3. BufferPoolInner — IoUringBufRing leaks on drop~~

IoUringBufRing internally stores its mmap region and backing buffers in ManuallyDrop fields. It requires an explicit release() call (which unregisters from io_uring) to free them. If a BufferPool is dropped without calling release_buffer_pool, the mmap and buffer Vecs leak.

Fix: Wrap buf_ring in ManuallyDrop inside BufferPoolInner and add a Drop impl that calls IoUringBufRing::drop() (the crate's provided unsafe cleanup function that frees memory without unregistering). Update into_inner() to take the ring out and mem::forget the wrapper to prevent double-free when the caller intends to call release().

[George-Miao]

Thanks for the PR.
For the 3 changes you mentioned:
1 looks good, go ahead. For 2 it was on purpose, since FrozenKey is used to unsafely sending non-Send data to other thread, if it's dropped, we can't run ThinCell's drop on other thread since that will cause UB. And for 3 we're planning for a major refactor for buffer pool, so please leave it as is.

[Berrysoft]

For 3, there might be ops still operating, referencing this bufing. Therefore, it's not safe to drop the buffers before releasing it.

[johnnyshields]

Reverted 2 and 3. Please give it a second look 🙇‍♂️