raylib 6.0 + linux arm64 builds

[Mantissagithub]

closes #320 (the "we implement raylib 6.0" + box2d arm build items)

adds linux arm64 builds (gh200 / gb200 / dgx spark / arm cloud boxes) by moving to raylib 6.0, which ships a linux_arm64 release.

changes

build.sh

• pick raylib-6.0_linux_amd64 / linux_arm64 from uname -m
• link -lX11 on linux, the 6.0 static lib calls X11 directly (undefined reference to XConvertSelection without it)
• web: 6.0 renamed the archive to libraylib.web.a
• impulse_wars: capnspacehook/box2d has no linux arm64 release (Extended Build capnspacehook/box2d#3 has been open ~a year), so on arm64 it builds box2d from source at the release tag df25d747 with the same flags as the fork's release.yml, including -DB2_MAX_WORLDS=65534
• -fsigned-char for clang and nvcc host code. char is unsigned on aarch64 linux and nmmo3 relies on -1 sentinels (see below). no-op on x86 and apple silicon
• nccl lib search uses /usr/lib/$(uname -m)-linux-gnu instead of hardcoded x86_64

raylib 6.0 api changes

• DrawCircleGradient takes a Vector2 center: boss_fight, checkers, go
• ModelAnimation.frameCount -> keyframeCount: tower_climb. 6.0 still samples gltf at 60fps and integer frames don't blend, so poses are unchanged

the nmmo3 char bug

flood_fill / cellular_automata store -1 in char buffers and check output[adr] == -1. with unsigned char those checks never match, biomes stay 255, and ALL_STONE[season] etc read out of bounds (asan: index 255 out of bounds for type 'char[4]'). terrain for a fixed seed (512x512, rng=42):

build	hash	tile types
x86_64	9d37398005190c20	12
aarch64, default char	adbdce1f0f3e2b27	3
aarch64, -fsigned-char	9d37398005190c20	12

also checked with -ffp-contract=off on both, same hash, so fma contraction on arm doesn't change the maps. building every env with -funsigned-char on x86 and diffing warnings only flagged nmmo3.

tested

• x86_64 (local, rtx 4060, cuda 13): --cpu builds 76/78 envs, same as raylib 5.5 before the change (matsci needs lammps, robot_arm is cuda only). native trainer builds for breakout, boss_fight, checkers, go, tower_climb, impulse_wars, nmmo3. breakout trains fine (~5.8M sps). --web builds breakout (emsdk 4.0.10)
• aarch64 + x86_64 (github ubuntu-24.04-arm / ubuntu-24.04): --cpu builds for every folder in ocean/, 82/85 on both archs with identical failures (matsci, robot_arm, osrs which is the shared dir for osrs_*). native trainer compiles and links on arm64 with nvcc 13 (NVCC_ARCH=sm_90, nvml stubs, no gpu on the runner) for breakout, nmmo3, impulse_wars, tower_climb. run: https://github.com/Mantissagithub/PufferLib/actions/runs/34910215978

not tested

• training on actual arm64 + gpu hardware, the ci runner has no gpu. if anyone on a gh200 / spark can run ./build.sh breakout && ./puffer train that closes the gap
• visually checking the boss_fight / checkers / go / tower_climb renders under 6.0, they compile but i haven't eyeballed them
• macos with raylib 6.0 (only the version string changed there)
• jetson is out of scope, jetpack has no nvml/nccl