feat(dflash2): add selector stop-gradient control

[maocheng23]

Summary

• add an opt-in training.dflash2_selector_stop_gradient control for DFlash2
• detach unary logits and draft hidden states only for selector CE, while preserving the primary DFlash/D-PACE/LK gradient path
• preserve PR support dflash2 #772 behavior by default and persist the resolved semantic in the resume contract
• document the option and cover config, resume metadata, selector parameters, and the backbone gradient boundary

Tests

• 54 targeted CPU tests passed
• 73 related DFlash/runtime tests passed on NVIDIA B200
• explicit CUDA backward smoke passed: selector gradients finite/nonzero and stopped backbone gradients exactly matched the primary-only reference
• Black check passed for all changed Python files
• git diff --check passed

Scope

This PR adds the selector stop-gradient control only. It does not change strict top-k candidate construction or add a gold-replacement selector curriculum.