[Shared Utils & Models] Implement tabular Q approximator with Numba acceleration

[saviornt]

Description

Create the core tabular Q-function implementation using NumPy arrays + Numba @njit for hot paths. Include Bellman update, ε-greedy action selection, and epsilon decay logic. Keep everything pure (no I/O) so it's reusable across edge and central.

Why: Tabular is the fastest MVP path for small discrete spaces and allows quick validation of the math.

Type

• Feature

Focus Area (pick one)

• Shared Utils & Models

Priority

• Critical

Acceptance Criteria

• TabularQ class with Numba-accelerated update and select_action methods
• Pure functions in core.py: update_q_value, select_action_greedy, decay_epsilon
• Exponential epsilon decay configurable via config
• Convergence test passes on a simple discrete toy env (e.g. 16-state grid world or frozen-lake-like)
• >90% test coverage for core logic and approximator
• Google-style docstrings with types, params, returns, and small examples
• MyPy strict passes

Blocker / Dependencies

• [Shared Utils & Models] Create Q-Learning config & types (Pydantic v2)

Notes / Links

• Target: shared/src/learning/q_learning/approximators/tabular.py
• Use @njit on update and selection; cache properties where sensible