Add body-free periodic GPU consumer for BlockLayout2D
blocks.cu: separate compilation entry with NX/NY = block window (halo
included), blockIdx.z selects the block; BlockHaloCopy refreshes halos
from owner cores, BlockStep pulls/collides core cells with the existing
compute_rho_u/compute_feq/collide_srt_no_force. Existing kernels and
refinement numerics unchanged.
block_stepper.py: PeriodicBlockStepper (D2Q9/SRT/FP32 double buffers,
doubly periodic layout); launch-limit and free-memory checks before any
allocation; uploads only halo pairs; spare buffer starts NaN; lifecycle
mirrors FixedPatchStepper.
Tests: CPU contract doubles (rejections before CUDA/allocation, initialize
guards, failure/close/retry) and one GPU case: 32x24, 8 steps, 1x1/2x2/4x2
splits bit-identical, FP64 oracle max abs <= 3e-6, drift reported.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
