fix(FifteenPuzzle): guarantee solvable boards; add optional difficulty tiers - #196

Problem: FifteenPuzzleEnv._generate_board shuffles all 16 tiles with a uniform random.shuffle. Only half of the permutations of a sliding puzzle are actually solvable, so ~50% of FifteenPuzzle-v0 episodes are impossible to solve — the agent can never reach the goal state and always runs out to the turn limit. Measured on main: 9983/20000 ≈ 49.9% solvable.

Fix: after shuffling, check solvability via the standard inversion / blank-row parity rule; if the layout is unsolvable, swap two non-blank tiles. A single transposition flips the inversion parity (and doesn't move the blank), which converts an unsolvable board into a solvable one. O(1), no distribution bias beyond the parity correction, and the goal state / tile set are unchanged.

Verification (20,000 generated boards):

• unsolvable: 0 (was ~50%)
• trivially already-solved: 0
• tile integrity (1–15 + one blank) preserved on every board
• ta.make("FifteenPuzzle-v0") → reset → step still works end-to-end

Additive difficulty tiers

Layered on top of the solvability fix, this PR also adds an optional difficulty argument ("easy" / "medium" / "hard") to FifteenPuzzleEnv, and registers FifteenPuzzle-v0-easy, FifteenPuzzle-v0-medium, and FifteenPuzzle-v0-hard.

The difficulty-tier idea comes from @MadBonzz in #169. This reimplements it additively so it composes with the solvable-board generator and doesn't change any existing behaviour:

• Non-breaking. The constructor keeps its max_turns argument; difficulty defaults to None, which is the existing full-shuffle-then-repair path. FifteenPuzzle-v0 is registered exactly as before.
• Solvable by construction. When a difficulty is set, the board is scrambled by a bounded number of legal blank slides from the solved layout (easy 1–5, medium 6–10, hard 11–20), so it is always solvable and its distance from solved scales with the tier — no parity repair needed. The scramble avoids immediately undoing its previous slide so the moves don't cancel out.
• Invalid difficulty values raise ValueError.

Verification (300 boards per tier + default):

• solvable: 300/300 for every tier and the default path
• mean misplaced tiles scales with difficulty: easy ≈ 3.9, medium ≈ 7.7, hard ≈ 10.5, default (full random) ≈ 15.0
• ta.make(...) → reset → step works end-to-end for the base id and all three variants
• FifteenPuzzleEnv(difficulty="…invalid…") raises ValueError

The core solvability change is unchanged from the original scope; the difficulty tiers are purely additive on top of it (new optional kwarg + new registered ids, no change to the prompt, rendering, wrappers, or the existing FifteenPuzzle-v0 registration).