Add Jinja2 backward kernel and PT2 wrapper templates for nobag embedding

[aagalleg]

This PR adds the Jinja2 templates that generate the SYCL backward pass
kernels and PyTorch 2.0 compilation wrappers for no-bag (sequence)
unweighted embedding lookups on Intel XPU devices. Together with the
forward templates, this completes the full training pass template set.

Depends on #82

Changes

Backward Kernel Templates (src/codegen/training/backward/)

• embedding_backward_split_kernel_templates.h: Jinja2 template
  generating both backward kernel variants for dense and split
  (rowwise Adagrad) paths:

  • SplitEmbeddingNobagBackwardCodegen{Dense|RowwiseAdagrad}UnweightedKernelWarpPerRow:
    warp-per-row kernel for short segments (SL < max_segment_length_per_warp),
    uses compute_grad_sum_unweighted_nobag for sub-group cooperative
    gradient accumulation
  • SplitEmbeddingNobagBackwardCodegen{Dense|RowwiseAdagrad}UnweightedKernelCtaPerRow:
    CTA-per-row kernel for long segments, processes all gradient
    contributions for a single embedding row
  • store_grad_sum helper (dense only): writes the accumulated
    gradient vector to grad_dev_weights with compile-time unrolling
    when kUseVecBlocking = false
  • split_rowwise_adagrad_table_update_kernel (split only): applies
    rowwise Adagrad weight updates with optional stochastic rounding,
    weight decay, and max-norm gradient clipping

• embedding_backward_nobag_unweighted_host_template.cpp: Jinja2
  template generating the host-side backward dispatch function
  split_embedding_nobag_backward_codegen_{dense|rowwise_adagrad}_unweighted_exact_xpu.
  Three-step dispatch pipeline:

  1. SplitEmbeddingBackwardFindLongSegments kernel: classifies each
     run-length encoded segment as short or long
  2. CTA-per-row kernel for long segments (submitted first for
     pipelining)
  3. Warp-per-row kernel for short segments
     — Calls transpose_embedding_input for CSR→CSC index sorting before
     kernel dispatch. The split variant additionally handles momentum
     tensors, LXU cache lookups, and deterministic algorithm selection

PT2 Wrapper Template (src/codegen/training/pt2/)

• embedding_forward_nobag_unweighted_pt2_wrapper_template.cpp:
  Single Jinja2 template (controlled by is_forward) generating both
  the forward and backward PT2 compilation wrappers for the split path:
  • Forward wrapper: delegates to
    split_embedding_nobag_forward_unweighted_xpu via
    torch::Dispatcher schema lookup
  • Backward wrapper: delegates to
    split_embedding_nobag_backward_codegen_rowwise_adagrad_unweighted_exact_xpu
    with full rowwise Adagrad parameter passthrough
  • Both wrappers registered via TORCH_LIBRARY_IMPL under the XPU
    dispatch key

cc: @flezaalv