GPU-05 workgroup barrier: __sync_threads + Barrier.arrive_and_wait #26

Add a CUDA-style workgroup barrier on the @Gpu path so cooperative kernels can
synchronize within a local workgroup (same scope as CUDA __syncthreads).

Expose two call forms from cthreads.sync that lower to the same GLSL
barrier (no host locks / no GLSL spinlocks):

• __sync_threads(): CUDA-familiar
• Barrier.arrive_and_wait(): beginner-friendly; not Barrier(...) construction

Motivation

• Enable shared-memory tiles / reductions later (__sync_threads is the first step)
• Keep one import surface: from cthreads.sync import Barrier, __sync_threads
• Reuse the existing GPU CallPlugin pattern (AST name / Attribute match)

Scope (explicit)

• Workgroup / block local only (GLSL barrier() + memoryBarrierShared())
• Not grid-wide sync (that stays multi-dispatch / host join)
• Not Python mid-run observe
• CPU Barrier(parties).arrive_and_wait() unchanged (host mutex/CV barrier)
• @Gpu rejects Barrier(n) / Barrier() construction with a clear error

Implementation sketch

• Stub: cthreads.sync.__sync_threads
• Plugin: gpu/compiler/translation/plugins/sync_threads.py
• Wire GpuFlow.expr_stmt so call statements lower through CallPlugins
• Tests: both forms emit barrier(); constructor rejected