Expand deterministic SDPA edge-case coverage to remaining pipelines

[hwanseoc]

Problem

#328 added deterministic coverage for negative row maxima, fully masked rows/columns, and zero-length packed sequences. Important SDPA execution paths remain uncovered by those edge cases:

• FP8 and MXFP8
• backward with cumulative cu_seq_len_*
• the d=256 pipeline
• BF16 fully masked backward

These paths may use different kernels or numerical handling, so the existing FP16/BF16 coverage does not protect them from the same NaN/Inf regressions tracked by #262.

Proposed coverage

Add focused deterministic cases, without forming a Cartesian product:

• Add one representative FP8 edge case and one MXFP8 edge case for supported forward/backward paths.
• Add zero-length cumulative cu_seq_len_* backward coverage after plumbing cu_seq_len_* through the backward test harness.
• Add one d_qk=256 representative that exercises the alternate pipeline.
• Run the fully masked-row backward case with BF16 and verify finite O, dQ, dK, and dV; LSE = -inf remains expected for masked rows.

Reuse the existing SDPA execution/reference harness and deterministic tensor hooks introduced by #328. Add version or capability guards where a path is not supported.

Acceptance criteria

• Each listed pipeline has a deterministic regression case.
• Outputs and gradients are checked against the reference and explicitly checked for unexpected NaN/Inf.
• Expected masked-row statistics are validated separately.
• Coverage stays representative rather than parameterizing every edge case across every dtype and feature.

Related: #262, #328