[FEA] Jagged Arbitrary Masked Self Attention support

[JacoCheung]

mcore/te does not support jagged attention with Arbitrary Masked (It only supports causal). But in our scenario, this is a desirable feature. In our sid_gr example, we now use padded dense with arbitrary mask which is memory and compute intensive.

Describe the solution you'd like

We can either wait for FA3 with arbitrary mask or write our only implementation. And then adapt it into sid example.

[JacoCheung]

It's now ready in fa)

[JacoCheung]

We need to implement a JaggedTransformerBlock, incorporating the Jagged arbitrary FA. @z52527
We need to start with the target grouped attention first. The mask should be like: