update(ssm): add BF16 timestep precision regression coverage

[Jaeyeong-CHOI]

Current main contains the production correction through PR #2270: the shared compiled SSM preprocessing casts both dt and dt_bias to float32 before addition. This PR preserves the contributor's original diagnosis and commit history while adding the isolated regression coverage that remains absent from main.

Root cause and coverage

• The defect was in compute_dt_compiled, before backend dispatch, rather than in the CUDA recurrence body or the float32 reference. MLX selected BF16 for float32 + bfloat16, so the processed timestep was rounded before the float32 state update.
• The regression uses zero state, x = B = 1, C = D = 0, and exactly representable inputs, making every next-state element equal host-f32 softplus(-2.5) without depending on a second GPU oracle.
• Float32 and BF16 activations run across BF16-to-float32-to-BF16 bias transitions. Existing tolerances are unchanged.
• The test is split into a child module to keep the parent parity file below the 500-line workflow limit.

Validation

• Focused CUDA build at candidate a9696bd8cb3d0550d312ce5b9c561c544f9886e4: passed.
• Focused GB10 runtime under gpu-lock: 6 passed, 0 failed, 0 ignored; all six SSM parity cases executed in 9.65 seconds.
• Rust formatting and diff whitespace checks: passed.
• Full CUDA workspace gate under gpu-lock: 12,247 passed, 0 failed, and 389 ignored across 163 suites, including doc tests.
• Broad CUDA Clippy with all workspace targets and -D warnings: passed.

PR #2270's preserved GB10 evidence reduced Granite state normalized RMS/max from 2.532402801e-3 / 4.127457397e-2 to 2.838613609e-8 / 1.701009653e-6 without changing tolerances, and its deterministic Granite fused-kernel versus graph comparison matched all 64 generated token IDs. This final PR does not claim a second production correction.

Bilingual technical reports are included under TECHNICAL_REPORTS/.

Closes #2142