adding model-only restoration to RL path

[azrael417]

This MR adds/changes the following things:

Weights-only model restore for RL (fine-tuning support)

Adds a coarse "load just the network weights" path for RL systems, mirroring the supervised load_model / load_checkpoint split, so a pretrained model can seed a new training run (e.g. modified reward or new
data) without dragging along the old optimizer state, replay/rollout buffer, normalizers, or step counters.

New API

• torchfort_rl_off_policy_load_model(name, checkpoint_dir) — TD3 / DDPG / SAC
• torchfort_rl_on_policy_load_model(name, checkpoint_dir) — PPO
• C (torchfort_rl.h) + Fortran (torchfort_m.F90) bindings, plus docs in c_api.rst, f_api.rst, and a "Fine-Tuning / Transfer Learning" section in usage.rst.

Semantics

• Restores only the online policy/critic (PPO: actor-critic) weights from a saved checkpoint; optimizers, schedulers, buffers, normalizers and counters stay freshly initialized.
• Off-policy target networks are re-initialized from the loaded online networks (θ′ ← θ), matching the standard init convention — not restored from the checkpoint. (load_checkpoint still restores everything
  verbatim for faithful resume.)

Tests

• New tests/rl/test_checkpoint.cpp (test_checkpoint_rl target) with round-trip tests for all four algorithms, asserting weights match after both load_checkpoint and load_model, and that is_ready is true after
  a full restore but false after weights-only (buffer not restored) — the first RL checkpoint round-trip coverage in the repo.

Test infrastructure (future-proofing)

• Config files are now resolved relative to the test executable (get_config_path in test_utils.h) instead of the current working directory, plus a build-tree copy of configs/, so RL tests run from any
  directory / either tree.
• test_utils.h made self-contained.

Cleanup

• Replaced the deprecated torch::optim::Optimizer::parameters() accessor (libtorch W603 warning) with a shared reset_optimizer_parameters() helper across all 9 RL + supervised load sites; the helper asserts
  the single-param-group assumption via THROW_NOT_SUPPORTED.

[azrael417]

/build_and_test

[github-actions]

🚀 Build workflow triggered! View run

[azrael417]

/build_and_test

[github-actions]

🚀 Build workflow triggered! View run

[github-actions]

✅ Build workflow passed! View run

[github-actions]

✅ Build workflow passed! View run

[romerojosh]

Looks mostly good to me. Just left a small comment about the added tests.

[romerojosh]

Another thing that came to mind here is that you'll want to add your new test_checkpoint binary to https://github.com/NVIDIA/TorchFort/blob/tkurth/inference-fixes/.github/scripts/run_ci_tests.sh. Annoyingly, this won't impact the workflow until this change is merged into master but we will want it in there.

[romerojosh]

Updates look good. You still need to add test_checkpoint_rl to .github/scripts/run_ci_tests.sh so this test gets run in the CI, but otherwise, we are good to go.