Fix NVIDIA workload startup settings and Docker readiness

[khluu]

Fix two NVIDIA workload startup settings and stop holding GPU hosts after a Docker server has exited.

This PR was authored with assistance from OpenAI Codex (GPT-6).

Changes

• Cap GLM-5.3 on 8x H200 at 65,536 tokens. In build #567, its default 1,048,576-token context required 53.45 GiB of KV cache with only 18.94 GiB available. The cap covers the configured 8K-input/1K-output benchmark and GSM8K sequence length. It changes the server context limit; full BFCL behavior under the cap still needs GPU validation.
• Replace DeepSeek-V4-Flash B200's removed use_fp4_indexer_cache=True flag with --attention-config.indexer_kv_dtype=mxfp4, preserving the old setting's meaning. Build #568 reached a GPU host but failed argument validation on the old flag.
• Track the Docker server container during readiness checks and return failure when it stops or disappears under --rm. In #567, GLM's engine failed around 13:40 UTC but readiness held the host until 14:20 UTC. Keep the native-process check, reset runtime tracking at startup, and try to print the last container logs on failure.
• Add regression coverage for stopped/removed containers, a running container becoming healthy, and native readiness without Docker. Run the lifecycle suite in the CPU pipeline-generation step and document readiness behavior.

Validation

• Server lifecycle tests: 7/7 passed (Docker operations mocked; existing health test uses localhost HTTP).
• Pipeline generation: 21/21 passed.
• Benchmark repetitions: 8/8 passed.
• Ingestion authentication: 2/2 passed.
• Parser smoke checks passed for both modified workloads with a stubbed lm-eval task registry.
• shellcheck lib/server.sh, shell syntax checks, git diff --check, and Buildkite's full online pipeline-schema validation passed.
• GPU performance, GSM8K, and BFCL runs have not been launched for this PR.

Remaining failures from #567

This addresses perf-eval configuration and lifecycle behavior. Completing NVIDIA recovery also requires:

• Capacity for DeepSeek's single-node 8-GPU request; #567's pod expired while unschedulable.
• An image containing vLLM #56825 for GLM-5.3-Flash's CUDA graph/event startup crash.
• The MSA compatibility revision from MSA #12, integrated by the still-open vLLM #56545, for MiniMax-M3's missing QuACK symbol.

AMD workload definitions and GPU scheduling policy are unchanged.