Add FlashAttention startup fallback

[mindbomber]

Fixes #31

Summary

• Add --attn_implementation to functioncall.py and jsonmode.py so users can run with eager or sdpa when FlashAttention is incompatible.
• Wrap FlashAttention model-load failures with an actionable message that explains the PyTorch/CUDA wheel mismatch and the reinstall workaround.
• Lazily import the heavy runtime dependencies so --help works before the full GPU/model stack is installed.
• Document the FlashAttention startup failure and fallback command in the README.

Verification

• python -m py_compile functioncall.py jsonmode.py -> passed
• python functioncall.py --help -> passed
• python jsonmode.py --help -> passed
• python -m ruff check functioncall.py jsonmode.py -> passed
• python -m ruff format --check functioncall.py jsonmode.py -> passed
• git diff --check -> passed
• AANA code-change guardrail -> pass / accept, AIx 1.0, no violations