Fix DSA top-k pinned buffer growth

[jayhenry]

Summary

DSA top-k offload used a monotonically increasing context ID for pinned storage, creating new pinned buffers every training step instead of reusing them.

This change:

• assigns stable slots to concurrently active microbatches;
• uses one (offload_slot, source_layer_idx) key for both the live SwapTensor and its reusable pinned storage;
• rejects slot reuse while the previous runtime entry is still active;
• propagates the slot through main and MTP sequence contexts;
• removes the obsolete context ID counter;
• adds multi-step micro1 and EP2/micro2 regression coverage.

Validation

• pytest -q tests/model/test_glm52_mtp_checkpoint_repro.py: 2 passed
• pytest -q tests/module/attention/test_dsa_mla.py::TestDSAAttention: 3 passed
• pre-commit hooks: passed, including ruff, mypy, and pydantic checks

Result

Each slot/source pair has one live offload lifecycle and one reusable pinned buffer, so pinned memory remains bounded across optimizer steps.