Fix leaks caused by mapped buffers

[teoxoy]

There are currently 2 instances where we leak:

• create a mapped_at_creation buffer without BufferUsages::MAP_WRITE, unmap it, make sure to not call .submit()
  • this will put the internal staging buffer in pending writes which won't get cleared because we don't call submit
  • brought up in buffer object leaking memory #5918 & Buffers Leaked when Device Handle is Dropped Before Buffers #5529 (comment)
• create a buffer with BufferUsages::MAP_WRITE, map it, unmap it, make sure to not call .maintain()
  • this puts the buffer in LifetimeTracker.mapped won't get cleared because we don't call maintain
  • this might not be a problem since users of the wgpu-core API should call maintain behind-the-scenes

[teoxoy]

I talked with @jimblandy about this and we concluded that we don't need to address these leaks because they only affect programs that never actually submit work to the GPU/don't poll devices.

The fixes are also not trivial and will most likely have perf impact.