Honor native attention backend selection in direct SDPA calls

[bghira]

Selecting native-math previously changed only Diffusers' dispatch registry. SDXL's AttnProcessor2_0 calls PyTorch SDPA directly, so it could still select flash attention. Scope PyTorch's SDPA selection alongside the existing Diffusers context for explicit native backends, restoring both when the backend changes.

The real CPU profiler regression previously observed aten::_scaled_dot_product_flash_attention_for_cpu and its backward despite selecting native-math. It now observes aten::_scaled_dot_product_attention_math with finite gradients. Additional tests cover backend switching and train/evaluation restoration for all explicit native SDPA choices.

Validation: .venv/bin/python -m unittest -v -f tests.test_attention_backend — 34 passed. Both the processor dispatch and backend-switch regressions failed before the fix and passed afterward.

Supplemental investigation: a tiny SDXL-shaped UNet with BF16 LoRA, math attention, gradient checkpointing, and SimpleTuner's AdamWBF16 completed three steps each on CPU and MPS. Convolution hooks observed recomputation during the first backward and every subsequent backward. This checks the control flow at reduced scale; it is not a full SDXL or RDNA2 hardware reproduction.

Related to #3195. This corrects the ineffective native-math diagnostic; resolution of the reported RDNA2 SIGSEGV requires testing on affected hardware.