feat: 🎸 add vit_tp_exp: DeiT-S tensor-parallel example, cross-checked…

[zengarden]

… L1-L4 against official DeiT

• Port the DeiT-S model verbatim from timm 0.3.2 plus the facebookresearch/deit recipe (train/eval loops, transforms, RASampler, MetricLogger) onto the tinyexp Cfg/Exp layout; attn.qkv split (D1) with bidirectional weight converters, checkpoints keep the official fused-qkv layout and load official DeiT checkpoints directly.
• TPAccelerator on torch.distributed.tensor.parallel: Megatron-style colwise/rowwise plan sharding attention on head boundaries; numerical equivalence vs single-rank asserted in CI (gloo), nccl+AMP on GPU.
• Full official recipe wiring (300ep AdamW/cosine, mixup+cutmix, repeated augmentation, EMA off at tp>1), mode=bench throughput/memory check, per-epoch checkpoints with resume, optional Redis raw-byte cache for slow shared filesystems (D7, bit-identical samples).
• Cross-check record in docs/vit_tp.md: L1 numerics (CI allclose); L2 official-ckpt eval 79.82/79.81 top-1 at tp=1/tp=2 (official 79.8); L3 full 300-epoch training from scratch on 8xH200 (rjob + redis cache): 79.84 top-1 / 94.99 top-5 in 17h56m (paper 79.8 / ~95.0); L4 TP-vs-DDP bench table.
• Engine hardening, each with a regression test: DDPAccelerator.reduce_sum stages cpu tensors through the device for nccl; store_and_run_exp resolves the canonical importable twin of a python -m main exp class so ray ships it by reference.