Chore/mypy policies rebased

[akacmazz]

What this does

This PR makes the lerobot.policies module 100% MyPy compliant by enabling strict type checking and adding comprehensive type annotations to all policy implementations.

Addresses #1719 ; fixes #1720 . (🔧 Enhancement)

Changes:

MyPy Configuration (pyproject.toml):

• Enabled strict MyPy checking for all 8 policy modeling modules
• Configured gradual typing approach with module-specific overrides
• Added disallow_subclassing_any = false for nn.Module compatibility

Base Classes (src/lerobot/policies/pretrained.py):

• Fixed ClassVar annotations for config_class and name attributes
• Updated get_optim_params() return type to dict | list[dict] for flexibility

Policy Implementations (50+ type hints added across 8 files):

• ACT (11 type hints): Return type annotations for __init__, reset, helper methods
• Diffusion (13 type hints): Fixed **kwargs typing, added parameter types
• TDMPC (8 type hints): Return types for __init__, reset, update methods
• SAC (8 type hints): Return types for main class and helper classes
• Pi0 (5 type hints): Return types for policy and helper classes
• Pi05 (5 type hints): Return types for policy and helper classes
• SmolVLA (3 type hints): Return types for SmolVLAPolicy
• VQBeT (6 type hints): Return types for policy and helper classes

Type Checking Fixes:

• Fixed None type narrowing with assertions
• Added 12 strategic type: ignore comments with TODO(Ensure the policy module passes MyPy type checks #1720) explanations for genuine edge cases:
  • **kwargs unpacking limitations (6 instances in Diffusion)
  • Literal type mismatches with config strings (2 instances in Pi0/Pi05)
  • Intentional signature overrides (1 instance in SAC)
  • torch.compile method assignment (1 instance in Pi0)
  • getattr dynamic type resolution (1 instance in SAC)
  • tuple unpacking with * operator (1 instance in SmolVLA)

How it was tested

1. MyPy Type Checking (Primary Test):

mypy src/lerobot/policies --config-file=pyproject.toml
# Success: no issues found in 49 source files ✅

Result: 131 MyPy errors → 0 errors 

2. Python Syntax Validation:
python3 -m compileall src/lerobot/policies/ -q
# No compilation errors ✅

3. Type Ignore Documentation:
grep -B1 "type: ignore" src/lerobot/policies/*/modeling_*.py | grep "TODO"
# All 12 instances have TODO(#1720) explanations ✅

4. Compatibility:
- Successfully rebased onto latest main (includes GR00T model from PR #1452)
- All new code passes MyPy checks including newly added policies
- Changes are type-annotation-only; no functional code modifications

How to checkout & try? (for the reviewer)

1. Verify MyPy passes on the policies module:
mypy src/lerobot/policies --config-file=pyproject.toml
Expected output: Success: no issues found in 49 source files

2. Check specific policy files:
mypy src/lerobot/policies/act/modeling_act.py --config-file=pyproject.toml
mypy src/lerobot/policies/diffusion/modeling_diffusion.py --config-file=pyproject.toml
mypy src/lerobot/policies/sac/modeling_sac.py --config-file=pyproject.toml

3. Verify all type: ignore comments have explanations:
grep -B1 "type: ignore" src/lerobot/policies/*/modeling_*.py | grep "TODO"
All 12 instances should show TODO(#1720) with explanations.

4. Review key changes:
- src/lerobot/policies/pretrained.py:49-50 - ClassVar annotations
- src/lerobot/policies/pretrained.py:160 - get_optim_params return type
- pyproject.toml:346-369 - MyPy configuration for policies module

5. Verify Python compilation:
python3 -m compileall src/lerobot/policies/ -q