feat!: complete embedding PyTorch Python API

[voltjia]

Summary

• Add the complete PyTorch Python embedding parameter list while keeping input, weight as the leading tensors.
• Keep the short default interface and retain the previous full interface as a [[deprecated]] compatibility overload.
• Implement max_norm/norm_type renormalization for shared CUDA backends and Ascend, and extend the existing test function.

Motivation

#808 restored the two leading tensors but intentionally omitted max_norm and norm_type. InfiniOps should expose the complete public Python contract even though it is a C++ library. This PR adds the missing behavior without multiplying every Python default into another overload.

This is a follow-up to #808; no standalone issue is associated with it.

Type of Change

• feat - new feature / new operator / new platform
• fix - bug fix
• perf - performance improvement (no behavioral change)
• refactor - code restructuring without behavior change
• test - adding or fixing tests only
• docs - documentation only
• build / ci - build system or CI configuration
• chore - tooling, formatting, or other non-code changes
• Breaking change

Platforms Affected

• CPU (WITH_CPU)
• NVIDIA (WITH_NVIDIA)
• Iluvatar (WITH_ILUVATAR)
• MetaX (WITH_METAX)
• Cambricon (WITH_CAMBRICON)
• Moore (WITH_MOORE)
• Ascend (WITH_ASCEND)
• PyTorch C++ bindings (WITH_TORCH)
• Build system / CMake / CI
• Python bindings / user-facing API

Smoke Test Result

Host: ssh nvidia
Image: accelerator-dev/nvidia:latest
GPU: NVIDIA A100, sm80
PyTorch: 2.10.0a0+b4e4ee8

cmake ... -DWITH_CPU=OFF -DWITH_NVIDIA=ON -DWITH_TORCH=OFF \
  -DCMAKE_CUDA_ARCHITECTURES=80 -DINFINI_OPS_OPS=embedding
cmake --build build-nvidia --target ops -j 16
Build passed.

python -m pytest tests/test_generate_wrappers.py -q
13 passed

python -m pytest tests/test_embedding.py --devices nvidia -q
57 passed

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
NVIDIA	Yes	targeted build passed	57 cases passed, including legacy and max_norm paths
Iluvatar	Yes	Not run locally - no Iluvatar host available	Uses the changed shared CUDA implementation
MetaX	Yes	Not run locally - no MetaX host available	Uses the changed shared CUDA implementation
Cambricon	No	N/A - not affected	N/A - not affected
Moore	Yes	Not run locally - no Moore host available	Uses the changed shared CUDA implementation
Ascend	Yes	Not run locally - no Ascend host available	Uses aclnnEmbeddingRenorm before aclnnEmbedding

Benchmark / Performance Impact

The default and no-max_norm paths retain the existing lookup kernel. Supplying max_norm adds selected-row norm reduction and in-place weight renormalization, matching PyTorch semantics.

Notes for Reviewers

API alignment

Interface	Signature	Alignment basis
InfiniOps full	embedding(input, weight, padding_idx, max_norm, norm_type, scale_grad_by_freq, sparse, out)	Complete PyTorch Python order; explicit out remains last per CONTRIBUTING.md.
InfiniOps short	embedding(input, weight, out)	Preserves the existing default-use interface without adding prefix overloads.
PyTorch public API	F.embedding(input, weight, padding_idx=None, max_norm=None, norm_type=2.0, scale_grad_by_freq=False, sparse=False)	Pinned PyTorch source
Deprecated compatibility	embedding(input, weight, padding_idx, scale_grad_by_freq, sparse, out)	Preserves the #808 full interface for migration.

• max_norm mutates selected weight rows in place before lookup, following PyTorch native renorm and its CUDA implementation.
• Ascend uses aclnnEmbeddingRenorm before lookup.
• padding_idx, scale_grad_by_freq, and sparse affect backward semantics; InfiniOps has no embedding autograd/backward path, so the forward implementation accepts and stores them without changing lookup values.
• The virtual backend contract gains max_norm and norm_type, which is why the PR is marked feat!; the previous user-facing full call remains available with [[deprecated]].
• Ruff 0.15.22 and clang-format 21.1.8 checks passed.