ref: support causal mask mode in flash attention

[Connor1996]

Summary

• add flash attention/quantized matmul extension implementation for week 2 tasks
• wire attention path updates for causal-mask capable flash attention
• remove a leftover debug sentinel write (-233.0) from src/extensions_ref/src/flash_attention.metal

Testing

• pdm run test --week 2 --day 4 -- -k 'task_2_flash_attention_cpu_small and no_mask' ✅
• pdm run test --week 2 --day 4 -- -k task_3_flash_attention_gpu ❌
  • runtime error: [metal::Device] Unable to load kernel flash_attention
  • causal case also hits ValueError: Invalid type str received in array initialization. in src/tiny_llm/attention.py:149