[Perf] Pipeline-friendly shard task submission in CacheStore

[mag1c-h]

Purpose

Decouple shard-level backend task submission to enable pipelining between dispatch and transfer stages.

• Backend's Load() is async, multiple shards can be processed concurrently
• Transfer stage can now start H2D on completed shards while waiting for slower ones
• Reduces latency when shard I/O times are imbalanced

Modifications

• Submit each shard to backend independently (not batched)
• Push ShardTasks to running queue immediately
• Each ShardTask has its own backend task handle for independent Wait()

Test

Modify	TensorSize	ShardNumber	BlockNumber	Load 100% from backend	Load 100% from Cache
Before	64KB	64	1024	314ms	264ms
After	64KB	64	1024	274ms	262ms