TBE backward: consume precomputed index-preproc tensors

[gchalump]

Summary:
Everything except the forward for moving grad-independent index preprocessing
off the TBE backward critical path. Two halves, folded into one diff:

• Backend driver (consume): adds an optional preproc_tensors (Tensor[]?)
  argument to the TBE backward driver so it can accept precomputed
  index-preprocessing outputs (transpose_embedding_input + find_long_segments)
  and skip recomputing them inline. When the arg is None (the default / common
  path), the driver runs the inline preprocessing exactly as before -- no
  behavior change. Applied across the backward host/meta/cu templates, the PT2
  CPU/CUDA wrappers, and split_embeddings_utils.

• PT2 autograd (route): threads the optional preproc_tensors bundle through the
  PT2 training autograd backward into the backend backward op. When the forward
  emits the 12 index-preproc grad slots (grad_outputs[1..12]), backward packs
  them into preproc_tensors and forwards them to the driver's consume path;
  otherwise preproc_tensors stays nullopt and the driver recomputes the preproc
  inline exactly as before.

Non-breaking on its own: no forward output is emitted yet, so grad_outputs stays
size 1 until the forward-emit diff (top of stack) is applied. The backward
simply gains the ability to accept and route the extra grads when they later
appear. Builds on the add-op base (D114645411, below in stack).

Differential Revision: D113624507

[meta-codesync]

@gchalump has exported this pull request. If you are a Meta employee, you can view the originating Diff in D113624507.