feat(nvidia): add Argmax provider

[voltjia]

Summary

• Add a CUB-backed native NVIDIA provider for the canonical Argmax operator.
• Support flattened float32, float16, and bfloat16 input with an int64 scalar output.
• Add focused Python coverage against torch.argmax.

Motivation

InfiniCore's basic Llama greedy decoding path needs a canonical InfiniOps replacement for the legacy sampling operator when top_k=1. The existing Argmax schema is aligned with torch.argmax(input, dim=None, keepdim=False), but InfiniOps did not have a native NVIDIA provider for the flattened reduction path.

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 (requires a ! in the Conventional Commits prefix or a BREAKING CHANGE: footer)

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

$ cmake --build /tmp/standard-llama/InfiniOps-build-argmax --target ops -j 16
[79/80] Linking CUDA shared module src/ops.cpython-312-x86_64-linux-gnu.so

$ pytest -q tests/test_argmax.py --devices nvidia
Running 6 items in this shard
......                                                                   [100%]
6 passed in 0.39s

The test build used CMAKE_CUDA_ARCHITECTURES=80 on an NVIDIA A100-SXM4-80GB.

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
NVIDIA	Yes	Target and focused test passed	f32, f16, and bf16; native and PyTorch implementation slots
Iluvatar	No	N/A - not affected	N/A - not affected
MetaX	No	N/A - not affected	N/A - not affected
Cambricon	No	N/A - not affected	N/A - not affected
Moore	No	N/A - not affected	N/A - not affected
Ascend	No	N/A - not affected	N/A - not affected

InfiniCore / InfiniLM integration

The dependent InfiniCore branch was built against this provider and ran paged-attention TinyLlama inference in accelerator-dev/nvidia:latest:

Query: Hello
Response: Certainly!
total_time: 285.2 ms

The run used top_k=1 and preloaded a trap for the legacy infiniop* C operator API; the process exited successfully without triggering the trap.

Benchmark / Performance Impact

N/A. This adds the previously missing native provider; it is not a performance comparison.

Notes for Reviewers

• The public schema is unchanged. This PR adds one native implementation and no overloads.
• The NVIDIA provider currently accepts only dim=None, keepdim=false, contiguous non-empty input, and up to INT_MAX elements. These constraints cover the basic Llama greedy sampling path; other torch.argmax modes can be implemented separately.
• The dependent InfiniCore integration is tracked in InfiniCore #1483.