Fix generic eval paths after changing the working directory

[haofrank]

Relative launches such as bash benchmarks/sglang_mi355x.sh source InferenceX's benchmark_lib.sh with a relative path. Magpie then changes into the isolated eval output directory, so the library's deferred BASH_SOURCE lookups can no longer find the GSM8K task or lm_eval_sitecustomize.py. A real Hyperloom synthetic GPU run completed throughput but failed accuracy evaluation with this error.

Resolve each generic benchmark script's directory before sourcing its dependencies. Deferred eval lookups now remain absolute after the working-directory change, including paths containing spaces. The persisted accuracy workflow, runner commands, and native AgentX launcher are unchanged.

Validation covers relative and absolute entrypoints through the real persisted-eval wrapper, task/patch lookup after the directory change, and persisted accuracy artifacts. The regression fails against the original scripts. Locally, 169 relevant tests passed. CI passed on Python 3.10 and 3.12: 629 passed and 1 skipped on each; coverage is 90.16%. The real Crusoe MI355X GPU rerun passed at Hyperloom 0ef71137f with Magpie a3339dc: 320/320 synthetic Qwen3-0.6B throughput requests, 15,637.91443 output tokens/s, and GSM8K strict-match 43/64 (67.1875%) with accuracy enabled and Magpie exit code 0. Torch/SGLang versions were unchanged. This exercises BaselineExecutor, not the full LLM optimizer loop. Detailed results: Hyperloom validation comment.

Required by AMD-AGI/Hyperloom#1675. This is a post-release fix on top of v0.3.0; it does not move the existing tag.