Backport smolvla grpo API hooks

[kufupa]

Title

feat(policies): backport smolvla grpo api hooks

Summary / Motivation

This PR ports SmolVLA GRPO API support from a local, working lerobot runtime patch into the upstream huggingface/lerobot repository in a minimal, isolated change. It updates SmolVLAPolicy to expose distribution-parameter access for GRPO-style training paths and aligns call patterns with current policy module imports used in the runtime. This is needed to unblock GRPO experiments relying on probability-distribution parameters (_get_distr_params_chunk) without altering environment setup or external runtime behavior outside this policy surface.

Related issues

• Fixes / Closes: # (if any)
• Related: # (if any)

What changed

• Backported src/lerobot/policies/smolvla/modeling_smolvla.py from the local patched environment runtime into this repo:
  • Added _get_distr_params_chunk for GRPO distribution-parameter path.
  • Added distr_queue queue bookkeeping setup in SmolVLAPolicy.
  • Updated select_action flow to handle the new model sampling tuple shape (actions, _ = model.sample_actions(...)).
  • Updated internal imports in modeling_smolvla.py to use the current in-repo policy module paths.
• Added GRPO-specific API hooks required by the smolvla-grpo pipeline while keeping change scope limited to the policy implementation.

No breaking changes expected. No migration required beyond updating downstream GRPO callers to use the available _get_distr_params_chunk path when needed.

How was this tested (or how to run locally)

• Validation attempted in this environment:
  • pre-commit run -a (failed: pre-commit not installed in this environment).
  • python -m pytest attempts were blocked by missing ML dependencies (torch not installed in this VM), so full local test run could not complete.
• Repro instructions for reviewer:
  1. Clone fork and checkout branch:
     • git checkout pr/port-smolvla-grpo-env-patch
  2. Install project/deps per repository instructions.
  3. Run targeted tests for smolvla policy behavior:
     • python -m pytest -q tests/policies/smolvla/test_smolvla_rtc.py
  4. Run lint:
     • pre-commit install
     • pre-commit run -a

Checklist (required before merge)

• Linting/formatting run (pre-commit run -a)
• All tests pass locally (pytest)
• Documentation updated
• CI is green
• Community Review: I have reviewed another contributor's open PR and linked it here: # (insert PR number/link)

Reviewer notes

• Focus review on src/lerobot/policies/smolvla/modeling_smolvla.py, especially:
  • _get_distr_params_chunk
  • select_action return-shape handling
  • import path updates to policy modules
• Confirm no environment runtime files are changed and this PR includes only intended policy API backport changes.
• Diff scope is intentionally minimal: one file changed (src/lerobot/policies/smolvla/modeling_smolvla.py).