Update SGLang to v0.5.17 and rebase Miles

[nanjiangwill]

Summary

• pin one unified SGLang v0.5.17 runtime for all cookbook models, including Kimi K3
• pin the rebased Miles fully-async runtime and migrate the GLM-4.7/GLM-5.2 recipes to Session v2
• retain disk, CPU/RAM, and CPU/NVMe weight-update paths with bounded staging lifetime and NUMA-interleaved canonical host memory
• add trajectory affinity and load-aware routing for dynamic rollout fleets
• validate FP8, ModelOpt NVFP4, mixed NVFP4/BF16, and Blackwell MXFP4 updates
• propagate explicit client cancellations through SGLang's structured abort lifecycle

Validation

• Stitch: 110 tests passed; Ruff lint and format checks passed
• SGLang: pre-commit passed; focused weight-loading suite passed with 48 tests and 2 skips
• all 12 model/update-path rows in the root README passed checksum and generation correctness checks
• GLM-5.2 CPU/RAM on a 3.17 TB host: 116.4 s preparation, 3.30 s engine pause, exact post-update text, tokens, and logprobs
• GLM-4.7 five-step E2E: four 8-GPU H200 trainer nodes, 48 one-GPU rollout replicas, logprob absolute difference 0.0172–0.0185, deltas v1–v5, and fleet catch-up through v5

SGLang is pinned at 0c79627e857eec795298a372adadf649209cdf2f on modal-projects/sglang:stitch-sglang-v0.5.17.

Miles is pinned at f83a68b296ca6f42d04e77e2b42c070c1ab02c70 on modal-projects/miles:stitch-miles-fully-async-swe.