feat(iluvatar): add native canonical attention providers

[gongchensu]

Summary

• Add native Iluvatar implementation-index 0 providers for the canonical FlashAttnVarlenFunc and FlashAttnWithKvcache APIs, reusing the existing native paged prefill and decode kernels.
• Adapt canonical paged-cache layouts and cumulative KV sequence offsets without allocating temporary device tensors.
• Keep existing CUDA-compatible providers on their original per-sequence-length and launch configurations while applying an Iluvatar-only 128-thread launch cap to ReshapeAndCacheFlash.
• Fix the standalone Iluvatar CMake link configuration by removing the unavailable CUDA::cuda_driver imported target; the backend only uses the CUDA runtime and cuBLAS APIs.
• Add Iluvatar canonical-attention smoke selection and an independent PyTorch paged-decode reference.

Motivation

Modern InfiniLM dispatches paged prefill and decode through the canonical
InfiniOps FlashAttnVarlenFunc and FlashAttnWithKvcache APIs. Before this
change, Iluvatar only had the legacy _infinilm compatibility providers, so
the canonical calls had no active Iluvatar implementation.

The canonical varlen API supplies cumulative KV offsets such as
[0, 130, 430], while the reused native paged-prefill kernel expects one
length per sequence, such as [130, 300]. This change lets the kernel derive
adjacent differences directly, avoiding an additional device allocation and
conversion kernel during InfiniRT graph execution.

The Iluvatar CMake configuration also referenced CUDA::cuda_driver, which is
not provided by the CoreX FindCUDAToolkit package. InfiniOps does not use the
CUDA Driver API on this backend, so only CUDA::cudart and CUDA::cublas are
required.

Validated with static, paged-attention plus InfiniRT graph, and explicit
FlashAttention plus InfiniRT graph inference using the 9G-8B model on
Iluvatar BI-V150.

N/A - no linked issue.

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
• Hygon (WITH_HYGON) — shared ReshapeAndCacheFlash template
• Mars (WITH_MARS) — shared paged-attention and cache templates
• T-Head (WITH_THEAD) — shared ReshapeAndCacheFlash template

Smoke Test Result

paste smoke test output here

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
NVIDIA
Iluvatar
MetaX
Cambricon
Moore
Ascend

Full `pytest` output (optional)

paste here

Benchmark / Performance Impact

Notes for Reviewers