Integrate Vortex with official SGLang 0.5.12 plugins

[Rockyeast]

Summary

• replace the vendored SGLang 0.5.9 source-patch path with the standard sglang.srt.plugins entry point for official sglang==0.5.12.post1
• register scoped hooks for ServerArgs, runtime integration, KV-pool sizing and initialization, fused KV writes, and PD receive handling
• preserve the Vortex planner's batch_size <= 1024 contract when SGLang resolves the request-pool capacity
• validate the SGLang/Vortex runtime contract and add the narrowly scoped Transformers 5.6.0 FlashAttention compatibility fix
• align planner dtypes, per-layer context propagation, and SGLang 0.5.12 KV-pool interfaces
• make generated indexer/cache modules namespaced, content-addressed, and atomically published
• fix missing Transpose graph registration and update installation/launch documentation

Motivation

Vortex previously depended on a repository-local, modified SGLang 0.5.9 checkout and import-time monkey patches. Environments that install official SGLang could parse Vortex configuration without activating the complete sparse-attention runtime, while newer SGLang pool and scheduler interfaces no longer matched the vendored integration.

This change makes Vortex a normal SGLang plugin, keeps the compatibility surface inside vortex_torch, and fails explicitly when the expected runtime contract is incomplete.

Validation

• python3 -m unittest discover -v -s tests -p 'test_*.py' (5 passed)
• python3 -m compileall -q vortex_torch tests with a temporary bytecode cache
• shell syntax checks for install_vortex.sh, install_vortex_glm.sh, and examples/misc/server_launch.sh
• TOML parsing and git diff --check
• Modal L40S runtime matrix with SGLang 0.5.12.post1: eager generation, CUDA Graph, overlap scheduling, CUDA Graph + overlap, DP=2, and checkpoint-backed per-layer Parameter
• AstraFlow RaaS launch smoke through --vortex-config to successful token generation

The GPU validation is single-node and does not claim multi-node coverage, other GPU architectures, trained-compressor accuracy, or production latency/throughput guarantees.