feat(attn): add switchable flash-attn and flashinfer backends

[lesj0610]

Summary

• add switchable attention backend policy support (auto, flash_attn, flashinfer, sdpa)
• resolve the selected backend once in runtime/generator code and use it consistently for prefill/decode paths
• keep recurrent prefill/checkpoint handling compatible with backend switching
• limit this PR to backend/core wiring in attn.py, generator.py, job.py, config.py, and transformer.py

Notes

• Refreshed onto v0.0.24
• Model-specific MLA support stays in follow-up PRs
• held_text replacement-character handling is intentionally kept out of this PR and remains in generator: always reconstruct held text when replacement chars appear #163

Validation

• Qwen3.5 abliterated EXL3 load smoke on the refreshed branch
• short English generation smoke completed successfully

[lesj0610]

Refreshed onto v0.0.24. This branch now only carries the backend/core runtime wiring needed for switchable attention backends.