[Feature Request] Resource control by persistent threadblock or multi-stream specialization

[chengyupku]

Support a context-based syntax to partition all resources into different groups, for example, dividing all thread blocks into two groups, or partitioning GPU resources into two streams to perform different tasks (specialization).

• e.g.,

with T.group("compute"):
	...
	
with T.group("communication"):
	...

In addition, it should support inserting barrier operations to guarantee correct synchronization between these groups.

Proposal: can refer to the T.Persistent primitive in TileLang.