Pin standard.def deps via requirements-direct.txt (fixes #19)

[nevasini1]

Summary

• Switches containers/standard.def to load containers/requirements-direct.txt instead of installing an unpinned package list inline
• Same pattern already used by vllm_debug.def and opus_4_6_codex_5_3.def, so standard.def now picks up the same pinned inspect-ai==0.3.150, transformers==4.57.3, and huggingface-hub==0.36.0 that resolve cleanly together
• Pins flash-attn==2.8.3 (was flash_attn unpinned) to match the working images

Fixes #19. As @hrdkbhatnagar noted in that thread, the maintainer-recommended pin is inspect-ai==0.3.150, which is what requirements-direct.txt already provides.

Verification

All 20 packages from the previous inline install list are present in containers/requirements-direct.txt:

accelerate, boto3, bitsandbytes, datasets, evaluate, lm-eval, openai,
pandas, scikit-learn, shortuuid, tokenizers, transformers, trl, peft,
tiktoken, inspect-ai, matplotlib, certifi, ninja, packaging

Container behavior preserved otherwise:

• AI CLI tool versions for standard.def left unchanged (claude-code@2.0.55, codex@0.79.0, gemini-cli@0.18.4, opencode-ai@1.1.59)
• inspect_evals still cloned with --depth=1 (no pinned commit), matching prior behavior of this image
• %runscript, %environment, %labels, %help blocks untouched

Test plan

• bash containers/build_container.sh standard succeeds end-to-end
• In the built image, python -c "import inspect_ai, transformers, huggingface_hub; print(inspect_ai.__version__, transformers.__version__, huggingface_hub.__version__)" reports 0.3.150 / 4.57.3 / 0.36.0
• pip check inside the image reports no resolver conflicts

(Re-opened from a clean fork; previous PR #46 closed.)

🤖 Generated with Claude Code

[nevasini1]

@hrdkbhatnagar @max-andr would be great if you can review this!