User code can lead to infinite panics in `interact` in `deadpool-sqlite`

[SergioBenitez]

Here's the interact method, for reference:

https://github.com/bikeshedder/deadpool/blob/2c7a0b7aee915e775f8d751e0a73ec884bae6f71/sqlite/src/lib.rs#L136-L146

If f() panics, it will do so while the lock is held. This poisons the lock. Every attempt to interact() thereafter will panic!() because of the unwrap() on line 144 (itself because of the unwrap() in 142). I would suggest the following changes:

1. Use a non-poisoning lock, like parking_lot::Mutex().
2. Propagate errors from spawn_blocking(), don't panic.

Alternatively, detect the error and close the connection so that the poisoned mutex is never reused.

[bikeshedder]

I wonder what's the best practice. My gut feeling is the following:

• The interact method should return a Result<InteractError, R>:

  /// This error is returned when the call to [`Connection::interact`]
  /// fails.
  #[derive(Debug)]
  pub enum InteractError {
      /// The provided callback panicked
      Panic(Box<dyn Any + Send + 'static>),
      /// The callback was aborted
      Aborted,
      /// rusqlite returned an error
      Rusqlite(rusqlite::Error),
  }

• Panics should can be propagated using this struct. The InteractError has a variant called Rusqlite so users can use ? inside the function and no double nesting takes place.
• When recycling connections detect the poisoned mutex and throw away that connection. Actually I feel like this is the correct approach anyways as that panic could've happened inside the rusqlite code and there is no safe way to check what caused the panic.

I just added this fix to the PR #115.