feat(linked): add MetaX FlashAttention providers

[voltjia]

Summary

• Add MetaX linked providers for the canonical flash_attn_varlen_func and flash_attn_with_kvcache operators using the installed flash-attn shared library.
• Register both providers in linked implementation slot 16 so builds can select them through the existing ops.json mechanism.

Motivation

InfiniCore's basic Llama/Qwen prefill and decode paths should call the canonical InfiniOps FlashAttention interfaces on MetaX. The installed MetaX flash-attn package exports an ABI with one additional optional extension tensor, so it needs a platform-specific linked adapter while preserving the canonical public operator contract.

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

MetaX, clean build using an ops.json selection:
-- Operator selection: /tmp/codex-ops-json-selection-20260819/InfiniOps/build/ops.json
-- Resolved 2 linked operator source(s).
-- Codegen torch op allowlist: argmax;index_select
[100%] Built target infiniops

The generated linked manifest contains only:

src/linked/torch/metax/ops/flash_attn_varlen_func/flash_attn.cc
src/linked/torch/metax/ops/flash_attn_with_kvcache/flash_attn.cc

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	Yes	full build passed	Qwen3-0.6B prefill and decode passed through the dependent InfiniCore branch
Cambricon	No	N/A - not affected
Moore	No	N/A - not affected
Ascend	No	N/A - not affected

Additional checks:

• python3 -m pytest tests/test_ops_config.py tests/test_generate_wrappers.py tests/test_resolve_linked_ops.py (84 passed)
• clang-format 21.1.8 --dry-run --Werror on all new C++ provider files
• git diff --check origin/master...HEAD

Benchmark / Performance Impact

N/A. This PR adds a MetaX adapter for an existing external implementation.

Notes for Reviewers

• Validated against flash-attn 2.6.3+metax3.2.1.3torch2.4. The MetaX ABI adds a trailing std::optional<at::Tensor>&; the adapter owns that extension argument and delegates through the canonical InfiniOps interface.
• The dependent build selects slot 16 for these two operators through ops.json, using the operator implementation selection added by feat(build): select operator implementations from ops.json #931. This PR adds no separate linked-provider selector or environment variable.