[Feat] Add Cache IO aggregation for Ascend

[NaganooMei]

Summary

• Add an Ascend Cache IO aggregation executor for shard-level H2D and D2H transfers.
• Add the Ascend shard IO aggregator and FFTS D2D dispatcher used by the aggregated executor.
• Wire CacheStore load and dump queues through the new executor abstraction while keeping the tensor executor as the default path.
• Add build and setup switches so upstream builds do not compile IO aggregation unless explicitly enabled.
• load path:

## Validation

Scenario	Baseline	With IO Aggregation	Result
GQA model, 32K IO load/dump	~4 GB/s per device × 8	~22 GB/s per device × 8	~5.75× bandwidth improvement
DeepSeek-V4 CacheStore IO	~3–4 GB/s per device × 8	~8–10 GB/s per device × 8	~2–3× bandwidth improvement
Default upstream build	IO aggregation disabled	Not compiled unless explicitly enabled	Default tensor executor path preserved
Code check	git diff --check upstream/develop..HEAD	Passed	No whitespace/check-format issues

[ygwpz]

💡 Suggestion: After aclrtSetDevice, the code should check for errors before proceeding. The current implementation checks this, which is good.

[ygwpz]

💡 Suggestion: The completed_ flag prevents adding new copies after launch. This is a good safety check. Consider also resetting this flag in Reset() to allow reuse of the dispatcher.

[ygwpz]

💡 Suggestion: The UnavailableAggregatedIOExecutor returns Status::InvalidParam for all methods when IO aggregation is not compiled. Since this is not a parameter validation error (the parameters may be perfectly valid), consider using Status::Unsupported() instead. This better communicates that the feature is simply not available in this build configuration.

The Status::Unsupported() method is available in ucm/shared/infra/status/status.h and maps to error code -50008, which the troubleshooting documentation describes as "Feature or operation not supported in current context" - exactly matching this scenario.