Add Craftax-Classic env to Ocean

[Infatoshi]

Single-header port of Craftax-Classic (Matthews et al., ICML 2024).
Follows the ocean/drmario template: craftax.h + binding.c +
config/ocean/craftax.ini.

Observation: 1345 float32 (7x9 tiles x 21 channels one-hot + 22 scalar) — matches the JAX Craftax-Classic-Symbolic-v1 layout exactly.
Action: 1 discrete in [0, 17) (noop, 4 moves, do, sleep, 4 place, 3 make-pick, 3 make-sword).
Log: perf = achievements/22, score, episode_return, episode_length, plus one counter per achievement.

Training sanity

RTX PRO 6000 Blackwell, ~30 seconds of training, default config from this PR:

metric	value
SPS	2.5M @ 99% GPU utilization
perf	0.69 (15/22 achievements unlocked)
episode_return	14.1
avg episode length	~312 steps

Agent cleanly learns wood → table → wood_pick → wood_sword → plant, progresses to drink / collect_stone / defeat_zombie. No NaN / divergence.

Implementation notes

• AVX-512 Perlin noise worldgen is enabled via a function-level __attribute__((target(...))), so the env itself requires no build.sh flag changes.
• Per-tile mob occupancy bitmaps give O(1) `has_mob_at` / collision queries.
• All game-logic helpers already operated on a single env state in the upstream craftax.c project, so no per-env vs batched rework was needed.

Commits

1. `ocean/craftax: Craftax-Classic env` — three files, additive, no core changes.
2. `build.sh: NCCL include/lib fallback via nvidia-nccl-cu12 wheel` — independent fix that mirrors the existing cuDNN fallback. Without it, builds fail on systems where NCCL is provided by the `nvidia-nccl-cu12` wheel in the active venv rather than a system install. Feel free to drop / split into its own PR if preferred.

Future work

Happy to follow up in separate PRs:

• Custom encoder that takes a compact uint8 observation and expands on GPU — Craftax's one-hot-per-tile structure means the dense Linear on 1345-float obs is mostly zero reads. Compact obs would cut PCIe traffic ~37x. Needs a way to plug a custom encoder / obs dtype per-env.
• `precision = bf16` flag in `[train]` for tensor-core-friendly matmul on Blackwell / Hopper. Our 3.0-era benches showed 1.7x on small-model minibatch microbenches; would want to re-measure inside 4.0's tuned CUDA trainer before committing to it.

Both are noted here rather than shipped so the env PR stays minimal.

Provenance

Derived from https://github.com/Infatoshi/craftax.c (standalone C/AVX-512 port, 47.8M SPS env-only bench). The Ocean port strips that project's batched-OpenMP harness since `src/vecenv.h` already OMPs across env instances.