[TF:XLA] Add partial implementation of tf.FIFOQueue for XLA devices (e.g., TPU).

The idea is to have a host-side queue of device tensors.

Operators dequeue_many, enqueue_many, and dequeue_up_to are not yet implemented because they require splitting/concatenating tensors, which will require calling into a compiled XLA compilation.

Refactor queue operator implementations into libraries separate from the kernel registrations.

Add support for ResourceOpKernels that are placed on non-CPU devices. Add support for allocating host-memory tensors during OpKernel construction.

PiperOrigin-RevId: 202590292