feat(mars): add native Mars kernels and htcc wrapper

[Ceng23333]

Summary

• Add Mars/HPCC device backend under src/native/cuda/mars/ (mirrors src/native/cuda/metax/ layout).
• Thin Operator<…, Device::Type::kMars> specializations reuse shared src/native/cuda/ops/*/kernel.h (same pattern as metax — not #include of metax headers).
• HPCC-only glue: hcblas, scripts/htcc_wrapper.sh, CMake WITH_MARS, htcc compile path.

Motivation

Enable InfiniOps on MetaX X203 / HPCC Mars (--device mars / InfiniLM --refactor-dev) without going through the MetaX/MACA stack. Layout policy: Mars impl stays under src/native/cuda/mars/, sharing generic CUDA ops like metax.

Closes # (none)

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

Note: Mars / HPCC (WITH_MARS) is the affected GPU backend; it is not yet a checkbox in this template.

Smoke Test Result

Host: MetaX X203 / HPCC, container infinilm-dev-refactor-dev (/opt/conda torch 2.8.0+mars, HPCC).
Branch feat/mars-native-kernels-htcc @ 9805393.
INFINI_RT_ROOT = /workspace/InfiniLM/build/integration/mars/prefix.

python -m pip install .[dev] --no-build-isolation --no-deps \
  --config-settings=cmake.define.INFINI_RT_ROOT=/workspace/InfiniLM/build/integration/mars/prefix \
  --config-settings=cmake.define.WITH_MARS=ON \
  --config-settings=cmake.define.INFINI_OPS_SMOKE_BUILD=ON
# Successfully built InfiniOps (cp310-linux_aarch64 wheel); install exit 0

python -m pytest tests -m smoke -q --devices cuda
# 92 passed, 35 skipped, 6235 deselected in 7.59s

Follow-up on this branch for smoke enablement:

• compile Mars torch unity with system g++ + -DUSE_HPCC=1 (MetaX/Moore pattern);
• TorchDeviceName<kMars> → "cuda";
• Mars torch C10 helper + generated/hand-written torch instantiations;
• htcc_wrapper.sh places .cu symlinks beside sources (not /tmp).

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
NVIDIA	no	N/A - not affected
Iluvatar	no	N/A - not affected
MetaX	no	N/A - not affected
Cambricon	no	N/A - not affected
Moore	no	N/A - not affected
Ascend	no	N/A - not affected
Mars/HPCC (WITH_MARS)	yes	pip smoke + pytest -m smoke passed (92 passed, 35 skipped)	Container infinilm-dev-refactor-dev; --devices cuda (Mars torch).

Full `pytest` output (optional)

s......ssssssssss....ssssssss........................................... [ 62%]
......................ssss.......s..........                             [100%]
92 passed, 35 skipped, 6235 deselected in 7.59s

Benchmark / Performance Impact

N/A

Notes for Reviewers

• Mars kernels live under src/native/cuda/mars/ and reuse src/native/cuda/ops like metax; do not rewrite to #include metax headers (Device::Type::kMars vs kMetax).
• Cherry-picked onto current master as a Mars-only change (does not include the older linked/ops.json commits from the Ceng refactor-dev line).
• Personal fork PR for the freeze-base review surface will be closed in favor of this upstream PR.
• Mars CONTRIBUTING smoke (pip install + pytest -m smoke) is green on this PR tip (9805393).

[Ceng23333]

Superseded by a stacked PR split that isolates Mars kernels from platform/torch:

Upstream (InfiniTensor/InfiniOps) — merge order

1. feat(mars): add WITH_MARS platform and htcc wrapper #941 — feat(mars): add WITH_MARS platform and htcc wrapper
2. feat(mars): add native Mars operator specializations #942 — feat(mars): add native Mars operator specializations (depends on feat(mars): add WITH_MARS platform and htcc wrapper #941)
3. fix(mars): enable torch bindings and CONTRIBUTING smoke #943 — fix(mars): enable torch bindings and CONTRIBUTING smoke (depends on feat(mars): add native Mars operator specializations #942)

Fork base-branch stack (isolated diffs)

1. Platform branch feat/mars-platform-htcc (upstream feat(mars): add WITH_MARS platform and htcc wrapper #941)
2. feat(mars): add native Mars operator specializations Ceng23333/InfiniOps#2 — kernels, base feat/mars-platform-htcc
3. fix(mars): enable torch bindings and CONTRIBUTING smoke Ceng23333/InfiniOps#3 — torch/smoke, base feat/mars-native-kernels

After #941 merges, retarget #942 → master; after #942 merges, retarget #943 → master.

Closing this monolithic PR (#939) in favor of the stack.