feat: add 2-head MLP skill generator skeleton with tests

[lisan-5]

Description

Implements the SubRep 2-head MLP generator skeleton.

This PR adds a new PyTorch SkillGenerator that takes an 8D observation and returns two outputs:

• payoff head: scalar output (shape (1,) for single input, (N, 1) for batch)
• motives head: 2D output [Safety, Fuel] (shape (2,) for single input, (N, 2) for batch)

Included in this PR:

• Shared MLP trunk (Linear -> ReLU -> Linear -> ReLU)
• Explicit weight initialization (Xavier uniform + zero bias)
• Input-shape validation with clear errors
• save(path) and load(path, map_location="cpu") model persistence methods (map_location optional, defaults to CPU)
• Generator README updates documenting architecture and I/O contract

Checklist

• Code follows project style guidelines
• Tests added/updated (tests/test_*.py)
• README updated (if interface changed)
• Verified locally (all tests pass)

Testing Notes

Run:

python -m pytest tests/test_generator.py -v 

To see example generator output printed from the save/load test:

python -m pytest tests/test_generator.py -v -s

Optional direct runtime check:

@'
import torch
from generator.skill_generator import SkillGenerator

model = SkillGenerator()
obs = torch.randn(8)
payoff, motives = model(obs)
print("payoff shape:", tuple(payoff.shape))
print("motives shape:", tuple(motives.shape))
print("payoff:", payoff.detach())
print("motives:", motives.detach())
'@ | python -