[Bugfix][MiniMax] Reset Lamport workspace after wake-up

[Ronald1995]

Purpose

Part of #53343. Split from #53344.

MiniMax fused RMS-QK uses a Lamport communication workspace containing flags, layout values, sentinels, and device pointers. Registered buffers can restore old bytes, but protocol state must instead be reset to a valid initial state after wake-up while preserving storage addresses captured by CUDA Graphs.

Changes

• expose the cached LamportWorkspace owner while retaining the tensor-returning compatibility API;
• register flag, layout, and workspace tensors as non-persistent buffers;
• rebuild Lamport sentinels, flags, layout, and pointer arrays in place;
• invoke the reset through the module post_weights_wake_up hook;
• add a focused protocol-state reset test.

Dependency

Depends on #53510 for the wake-hook dispatcher. #53510 itself depends on #53509.

This is a stacked PR. Merge order:

1. [Bugfix][MoE] Restore derived buffers after weight reload #53509
2. [Bugfix][Humming] Recover locks after sleep and reload #53510
3. this MiniMax PR

GitHub will reduce the diff as the prerequisite PRs merge.

Test Plan

pytest -q tests/kernels/core/test_minimax_reduce_rms.py -k lamport_workspace_reset

Changed files pass Ruff, git diff --check, and Python compilation. Distributed CUDA kernel validation is delegated to CI.

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @Ronald1995.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork