Add flash decode v2

[harsh-nod]

This PR at support for a new variant of flash decoding. In this version we use a split – K approach, where the first kernel compute the attention parallelizing over the sequence length, and the second kernel sums up all the results. In order to make this work, we had to add support for the following

• SymbolicAliases - these are language constructs that allow us to say that one symbol is associated with another symbol with a specific relationship. This is implemented as constructing additional work group constraints telling constraints and wave constraints. They are ignored during expansion.
• Added a template for flash decoding that is used in both the lit test as well as the end to end tests