Add Bomberman two-player environment (revives #93) - #198

Revives the two-player Bomberman environment, originally proposed by @romirthedev in #93. That PR targeted an older State/observation API and no longer imports against current main, and the author is no longer active, so this ports it forward and verifies it end-to-end.

What this is

A turn-based two-player Bomberman on a grid with indestructible (#) and destructible (+) walls. Each turn a player moves ([up]/[down]/[left]/[right]), waits ([stay]), or drops a bomb ([bomb]). Bombs detonate after a timer with a cross-shaped blast that destroys destructible walls and eliminates players caught in it. Last player standing wins; both caught in one blast, or reaching the round limit, is a draw. Registered as Bomberman-v0.

Changes vs. the original submission

• State is created in reset() as TwoPlayerState; reset(num_players, seed) signature.
• Observations use ObservationType; invalid moves use set_invalid_move(reason=).
• Single-winner outcomes use set_winner(); mutual destruction and the round limit resolve to a draw.
• Dropped the unused numpy dependency and the dead offline_renderer import; added get_board_str().
• Bug fixes beyond the API port: player spawn cells could land on an indestructible pillar (for even grid sizes) — the spawn cell is now forced clear; and dropping a bomb onto a cell that already holds a bomb is now rejected as an invalid move.
• The action regex already matched bracketed tokens, so it works cleanly under ActionFormattingWrapper.

Verification

• ta.make("Bomberman-v0") → reset / observation / valid move all work.
• Terminal branches verified with deterministic scenarios: bomb eliminates the opponent → single winner; both players in one blast → draw; round-limit → draw; repeated invalid actions → loss.
• Random-play smoke across 6 seeds: every episode terminates with no exceptions.

Closes #93 if you'd rather track it here; otherwise feel free to keep #93 open and land this instead.