[Feat] Add layerwise load-ahead window

[dante159753]

Summary

• Add configurable layerwise_load_ahead for vLLM layerwise KV loading.
• Keep the default at 1 to preserve the existing one-layer lookahead behavior.
• Refill the load window by local layer order after each layer wait, avoiding duplicate submissions.
• Document the tuning option in the example config and PipelineStore guide.

Validation

• Manually ran the new scheduling tests because local Python does not have pytest installed.
• Ran python -m compileall ucm\integration\vllm\ucm_connector.py ucm\integration\vllm\tests\test_layerwise_load_ahead.py.
• Commit hooks passed: codespell, black, isort.

Notes

layerwise_load_ahead=1 preserves the previous behavior. Values like 2 or 4 can improve overlap when layer load latency is higher than per-layer inference latency, but should be tuned with Cache Store queue, host buffer, and H2D stream pressure in mind.

[ygwpz]

Missing test coverage: no test for load_ahead > total_layer_count. What behavior is expected when the window exceeds available layers? Consider adding a test case for this edge condition.

[ygwpz]

Missing test for validation errors: _get_layerwise_load_ahead should raise ValueError for negative or non-integer values, but there's no test verifying this behavior.

[ygwpz]

The should_refill_window check adds to _waited_load_layers before popping from load_tasks. If the same layer is revisited (e.g., rollback or MTP paths), this could cause issues. Should the check happen after the pop?

[ygwpz]

The _submit_next_load_layers loop condition self._next_load_layer_index < len(self.layer_ids) stops silently when index exceeds. Should there be a debug log or assertion to clarify expected behavior?