Add Ascend Cache IO aggregation copy path

[NaganooMei]

Purpose

This PR adds an Ascend CacheStore IO aggregation path for small H2D/D2H shard copies.

The original CacheStore transfer path submits one async copy per tensor fragment. For workloads where one CacheStore shard contains many tensor fragments, this creates many small H2D/D2H submissions and increases CPU/runtime scheduling overhead.

This PR introduces an Ascend-only aggregation stream. It treats one CacheStore shard as one IO object, uses a staging buffer on device, and launches FFTS SDMA descriptors to scatter/gather between the staging buffer and device tensor addresses.

For PLATFORM=ascend / RUNTIME_ENVIRONMENT=ascend builds, CacheStore enables IO aggregation by default. When use_layerwise=true, the vLLM connector disables CacheStore IO aggregation to preserve the existing layerwise path behavior. Non-Ascend runtimes do not compile or allow this feature.

Modifications

• Add an explicit trans-layer factory for CacheStore IO aggregation:
  • IoAggregationStreamConfig
  • MakeIoAggregationStream(config)
• Keep CopyStream as the CacheStore-side stream owner, but let IO aggregation use one outer stream object with internal lanes.
• Add shard-level multi-fragment H2D/D2H stream APIs so LoadQueue and DumpQueue can submit one CacheStore shard as a scatter/gather object instead of looping over tensor fragments.
• Add AscendIoAggregationStream and AscendShardIOAggregator.
  • H2D load: host shard buffer -> device staging buffer -> device tensor fragments.
  • D2H dump: device tensor fragments -> device staging buffer -> host shard buffer.
  • Internal lane count comes from cache_stream_number.
  • Pipeline depth and max ready lanes keep default values and remain configurable for advanced tuning.
• Add FftsSdmaDispatcher to build and launch FFTS SDMA copy descriptors.
• Add Ascend runtime/build isolation:
  • Ascend builds enable UCM_RUNTIME_ASCEND_IO_AGGREGATION.
  • Other runtimes reject cache_io_aggregation=true.
• Add config validation for unsupported combinations:
  • IO aggregation requires Ascend runtime support.
  • IO aggregation is incompatible with use_gdr.
  • Invalid pipeline depth / ready-lane values fail early.
• Keep existing GDR, CUDA, simulation, and normal Ascend copy paths unchanged when IO aggregation is disabled.

Config Behavior

• cache_io_aggregation defaults to enabled in Ascend builds.
• cache_io_aggregation is forced to disabled by the vLLM connector when use_layerwise=true.
• cache_io_aggregation_pipeline_depth defaults to 2.
• cache_io_aggregation_max_ready_lanes defaults to 8.
• The example YAML does not expose IO aggregation tuning knobs; normal users only need to build with Ascend runtime support.

Test

• Verified Ascend build path with IO aggregation enabled.
• Verified default CacheStore load/dump behavior with IO aggregation disabled.
• Verified CacheStore load H2D path with IO aggregation enabled.
• Verified CacheStore dump D2H path with IO aggregation enabled.
• Verified unsupported runtime/config combinations fail during config validation.