Avoid blocking in Transaction::close() when there's a cancelled async write when possible

[tgoyne]

We can't cancel the wait for the write lock if another process currently holds it, but if the DB for the current Transaction holds it and we're just waiting in the write queue then we can safely close the Transaction immediately and simply skip over it when it would get its turn to write.

This used to require writing some pretty convoluted code to run into, but with structured concurrency it's much easier to accidentally write something that'll deadlock. For example, the following code does:

let realm = try Realm()
realm.beginWrite()
let task = Task {
    let realm = try Realm()
    try await realm.write {
        ...
    }
}
task.cancel()
try await task.value

[jedelbo]

When I run realm-object-store-tests "SharedRealm: async writes" the test sometimes fails with

terminate called after throwing an instance of 'std::bad_weak_ptr'
  what():  bad_weak_ptr

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
realm-object-store-tests is a Catch2 v3.0.1 host application.
Run with -? for options

-------------------------------------------------------------------------------
SharedRealm: async writes
  close()
  before async_begin_transaction() callback
-------------------------------------------------------------------------------

and sometimes with

terminate called after throwing an instance of 'std::runtime_error'
  what():  timed_sleeping_wait_for exceeded 30000 ms

when in "while another DB instance holds lock"

[tgoyne]

I can reproduce both of those by enabling the sleep in DB::AsyncCommitHelper::main().

[tgoyne]

Closing in favor of #6486.