[Attention][Core] Add generic prefix-anchored SWA on SM70

[yangzhuxinyzx]

Purpose

Replacement for #239, rebased onto current main. Adds prefix-anchored sliding-window attention as an explicit, default-off inference-engine capability. Admission is based only on hardware, operator, tensor/cache, graph, and concurrency contracts; it does not inspect model/checkpoint/architecture identity.

The engine option is AttentionConfig.prefix_anchored_decode_window. The prefix remains globally visible while generated-token KV outside the live window is masked and reclaimed. Unsupported combinations fail up front, and missing runtime metadata fails closed.

Key audit fixes over #239

• removed DeepSeek/OCR-specific classes and checkpoint config wiring
• moved activation to model-agnostic AttentionConfig and base Attention
• rebased the CUDA changes over the current Flash-V100 kernel stack
• rejected non-positive windows, non-fp16 KV, non-SM70, MLA/local/sink attention, CP, speculation, KV connectors/offload, and unsupported backends
• fixed the missing eager/piecewise metadata path
• made masked extension availability and metadata consistency hard startup/runtime gates
• preserved original-author attribution without carrying unsigned commits

Test Plan / Result

• pytest -q tests/config/test_prefix_anchored_swa.py tests/v1/attention/test_prefix_anchored_swa_mask.py tests/v1/core/test_single_type_kv_cache_manager.py: 32 passed
• SM70 source build: CUDA_HOME=/usr MAX_JOBS=4 TORCH_CUDA_ARCH_LIST=7.0 python setup.py build_ext --inplace: passed
• pytest -q tests/kernels/attention/test_sm70_flash_v100_anchored_swa.py on V100: 19 passed
• changed-file pre-commit suite, including Ruff, clang-format, mypy, SPDX and config/docs checks: passed
• same-toolchain SASS comparison to current origin/main: 55/55 decode and 64/64 paged-prefill default-off instances byte-identical

No model end-to-end or throughput claim is made. Refs #239.

[yangzhuxinyzx]

Audit note: the required pre-run-check passed. The repository-wide manual pre-commit job failed because the all-files formatter rewrote 228 pre-existing files; its rewritten-file set has zero overlap with this PR's 26 changed files. The complete changed-file pre-commit suite passes locally on the PR head. This known main-branch formatting drift is therefore not attributed to #281.