fix(dflash): gate flex_attention BACKEND=TRITON on torch>=2.11 (fixes ROCm)

[zhentaocc]

Problem

OnlineDFlashModel._forward_draft_blocks (shared by DFlash / Domino / DSpark)
pins the flex_attention kernel option {"BACKEND": "TRITON"} to keep every
DFlash-family batch on the general Triton Flex Attention kernel (AUTO can
otherwise route short-query batches to the flex-decoding kernel, whose config
set can be empty for DFlash's sparse BlockMask).

However, the "BACKEND" kernel option is only understood by the inductor
flex-attention lowering on torch >= 2.11, where it is sanitized out of the
generated Triton constexprs. On older builds the string leaks into the
generated kernel as a bare identifier and compilation fails with:

NameError: 'TRITON' is not defined

This affects any pre-2.11 torch, and in particular current torch ROCm
wheels, which ship a 2.9.x-based build.

Fix

Gate the kernel option on the torch version rather than on the vendor:

• torch >= 2.11 → {"BACKEND": "TRITON"} (unchanged CUDA behaviour)
• older → {"FORCE_USE_FLEX_ATTENTION": True}

FORCE_USE_FLEX_ATTENTION has been supported since torch 2.5 and selects the
exact same general Triton Flex Attention kernel — verified by dumping the
generated inductor code on torch 2.11 and confirming both options produce a
byte-identical triton_tem_fused_flex_attention kernel (no flex-decoding, no
FLASH/CuteDSL template). This fixes every pre-2.11 build, ROCm included, while
leaving the CUDA path identical.

test_online_wrapper_forces_standard_triton_flex_backend is updated to assert
the version-appropriate kernel options.