[PyTorch] Add torch_stable compatibility layer for incremental migration to the torch stable ABI

[pggPL]

Description

First step of the incremental migration of transformer_engine_torch to the torch stable ABI (so that one wheel can work across torch versions).

Key observation: pybind11 itself does not depend on the torch ABI (only on CPython) — the torch-ABI dependency comes from torch/extension.h / at::Tensor / torch's pybind casters. So the extension can migrate to torch::stable incrementally, file by file, while staying on pybind11.

This PR adds the foundation:

• csrc/torch_compat.h — a thin compatibility layer. Migrated code is written against the torch_compat surface, which is restricted to what torch::stable provides. In the default build it maps to the full torch ABI (torch_compat::Tensor is at::Tensor, torch's own pybind caster applies) — no stable headers are touched, so torch >= 2.1 keeps working unchanged. With NVTE_TORCH_STABLE_ABI=1 (experimental, requires torch >= 2.14) it maps to torch::stable (including torch::stable::tensor_from_pyobject / tensor_to_pyobject from Adding conversion from PyObject to torch::stable::tensor pytorch/pytorch#183323).
• csrc/extensions/stable_tensor_caster.h — a pybind11 type_caster<torch::stable::Tensor>, compiled only in stable mode; a torch.Tensor argument is unwrapped into a stable tensor sharing the same TensorImpl.
• makeTransformerEngineTensor + convertTorchShape overloads for stable tensors in common.cpp (stable mode only).
• First ported file: csrc/extensions/recipe.cpp (compute_amax, fused_amax_and_scale_update_after_reduction) now targets the torch_compat surface. No behavior change in the default build.
• Build flag plumbing in build_tools/pytorch.py (NVTE_TORCH_STABLE_ABI → -DNVTE_WITH_TORCH_STABLE -DTORCH_TARGET_VERSION=2.14).

Follow-ups: port the remaining csrc files in small PRs, growing torch_compat.h as needed (missing stable APIs: RNG/Philox, c10d, some ops — being upstreamed to pytorch separately), then flip the stable mode on for torch ≥ 2.14.

Type of change

• Documentation change (change only to the documentation, either a fix or a new content)
• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Infra/Build change
• Code refactoring

Changes

• Add torch_compat.h dual-mode compatibility layer (full ABI by default — torch >= 2.1 unaffected; stable ABI opt-in)
• Add pybind11 type_caster<torch::stable::Tensor> (stable mode only)
• Add NVTE_TORCH_STABLE_ABI build flag (experimental)
• Port extensions/recipe.cpp to the torch_compat surface

Checklist:

• I have read and followed the contributing guidelines
• The functionality is complete
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes