test: add SDPA gradient health tests (fixes #44928)

[Lemniscate-world]

Fixes #44928

Summary

Adds gradient health validation tests for scaled_dot_product_attention with dense attention masks. These tests catch gradient explosion bugs such as #44928, where 3D position_ids force SDPA Math fallback, causing BF16 collapse and NaN gradients in Qwen3.5 RLHF training.

Tests added

• test_sdpa_finite_gradient_standard: SDPA must produce finite gradients with standard inputs
• test_sdpa_finite_gradient_dense_mask: Dense causal mask must not cause NaN/exploding gradients
• test_sdpa_gradient_scaling_consistency: Gradient scales approximately linearly with query scaling

Motivation

Issue #44928 reports catastrophic gradient explosion (NaN) when using Qwen3.5 with 3D position_ids that force SDPA Math backend fallback under BF16. These tests establish baseline gradient health expectations and would have caught the regression.

Detection

These patterns are monitored by NeuralDBG, a causal inference engine for DL training, via its gradient health transition events.

[Lemniscate-world]

Detected via NeuralDBG gradient health monitoring. The test patterns here (finite check, dense mask safety, scaling consistency) mirror the gradient_health_transition events that NeuralDBG captures at runtime.

CC @ArthurZucker for visibility (related to SDPA attention).

[github-actions]

---

CI recap

Dashboard: View test results in Grafana
Latest run: 28593785482
Result: failure | Grafana metrics are not available yet.

Code quality check failed: test jobs were skipped. Fix the code quality issues and push again to run tests.