fix(vllm): support synchronous GPUDirect loads

[ketor]

Summary

• implement load_async=false for the vLLM connector
• execute external GPUDirect loads in start_load_kv, before the request forward pass
• keep the existing asynchronous WAITING_FOR_REMOTE_KVS path unchanged for load_async=true
• validate the setting as a JSON boolean and document when synchronous mode is required

Problem

The async path issues remote GPU writes from get_finished() after model compute has already launched. On hybrid state-cache models, vLLM does not currently expose a per-block ownership fence proving those destination blocks are disjoint from concurrent recurrent-state compute.

On GLM-5.3-Flash TP8, 64K-input C10 external-cache replay consistently corrupted in-flight CUDA state:

• C1 and C2: 20/20 successful
• C10 async with node dedup enabled: 35/100 successful, then illegal memory access
• C10 async with node dedup disabled: 21/100 successful, then CUDA/NVLink illegal access

The native client reported exact-length GET hits and zero I/O errors, so retrying transport or changing node dedup did not address the ownership race.

Validation

Hardware: xb01-0064, 8x B200. Model: GLM-5.3-Flash TP8+EP, no MTP. Workload: 100 prompts, 65,536 input tokens, 500 output tokens, concurrency 10.

With load_async=false:

• 100/100 successful
• 69,121 total tok/s, 523 output tok/s
• 80,488 dfkv GET hits / 57,465,604,224 bytes read
• zero GET misses and zero GET/PUT I/O errors
• 156K-token restart round-trip also completed through dfkv; the target response contained the requested verification code

Tests:

• worker lifecycle suite: 9/9 passed in the dedicated vLLM image