Nemotron: fix SSM backward thread limit on M1

[mrv777]

On M1 Max, tf_train_ssm_back reports a 768-thread limit, but training dispatches 1,024 threads per head. The launch can leave gradients unwritten without a command-buffer error.

Declare a 1,024-thread limit and add regressions for gradients, seeded checkpoints, shared groups and checkpoint boundaries, including the production 64-head/8-group shape. Arithmetic and dispatch sizes are unchanged.

Validated on M1 Max with Zig 0.17.0:

• Both regressions fail without the annotation and pass with it. The tests cover 38 GPU cases with FP64 references and guards, plus finite-difference gradient checks.
• Full suite: 1,292 passed, one known GLM sparse-attention failure (GLM: fix sparse attention thread limit on M1 #704), two Linux-only skips.
• With GLM: fix sparse attention thread limit on M1 #704/GLM: fix KDA and router thread limits on M1 #708 applied temporarily: 1,297 passed, zero failed, two skips; all 145 build steps pass.
• Native build and server goldens pass (zero unexpected differences). Formatting passes; zero new lean findings (534 inherited).

This PR is independent of the GLM fixes. Raw logs remain local. Full-model training and other GPUs remain untested.

[ashhart]

Thank you, @mrv777. Gradients left unwritten with no command-buffer error would have been hard to find from a training run, and the new tests pin every shape that matters. Your commit landed in 1.0.9 under your name (5d5e0bda0), and the PR's head is merged into the history with no change to the tree (c41fdf1ac), so GitHub shows this PR as merged.

Your gradient checks pass on an M3 Ultra and an M5 Ultra (2 of 2 on each, with the same errors on both chips), and the macOS host suite passes. We haven't timed a Living Weights training step with it; the arithmetic and dispatch are unchanged, so we expect no difference. M1 is your receipt.