Add experimental Lantern-backed Tensor API

[dfalbel]

What

Adds a torch::experimental::Tensor C++ facade backed by Lantern calls. It supports direct use in Rcpp exports, including fluent code such as input.relu().reshape({-1}).

The initial handwritten facade covers core Tensor behavior and conversions, while generated methods expose the broader Tensor namespace from src/gen-namespace.cpp. The generator currently emits 758 overloads and audits coverage of 815 public Lantern Tensor entry points.

Why

Packages using torch currently receive handle types but cannot write a LibTorch-style Tensor API without linking directly against LibTorch. This facade keeps the existing Lantern ABI boundary while providing a familiar C++ interface.

Impact

• Adds an opt-in API under torch::experimental; existing APIs are unchanged.
• Resolves Lantern symbols through registered R C-callables.
• Adds representative runtime tests for Tensor input/output, chaining, generated methods, integer arguments, and boolean returns.
• Adds a freshness and coverage test for generated methods.

Validation

• python3 tools/generate-experimental-tensor.py --check — passes (758 generated overloads; 815 Lantern Tensor entry points covered).
• testthat::test_file("tests/testthat/test-experimental-cpp-codegen.R") — passes.
• testthat::test_file("tests/testthat/test-experimental-cpp-api.R") — passes all 5 runtime assertions against the installed branch and configured Lantern runtime.
• The package was installed from this branch with BUILD_LANTERN=1 R CMD INSTALL . before running the tests.

[dfalbel]

Added the experimental tensor creation API in e6ae16dc5.

Exposed under torch::experimental:

• empty, empty_strided, empty_like
• zeros, zeros_like, ones, ones_like
• rand, rand_like, randn, randn_like
• randint, randint_like, randperm
• arange, linspace, logspace
• eye, full, full_like, scalar_tensor

Each factory supports LibTorch-style default options, and applicable overloads accept the existing torch::TensorOptions handle. Compile coverage exercises every operation. The focused suite passes all 10 runtime assertions, and the generator audit still passes with 758 generated Tensor methods and 815 Lantern Tensor entry points.