Support causal flash attention

[jopperm]

This PR adds support for causal FA:

• Keeps encoding on row-vector tensor operations, as must be left untouched when lowering to the SIMT program.
• Extends the pattern matching helper that determines whether a tensor is transposed, to look through advance operations. (The second attention loop uses a transposed tensor pointer that is tt.advance'd between the loops.)

[jopperm]

For D_HEAD=128, we're getting tt.make_range ops that are smaller than the subgroup size; I don't know how to lower these yet. Checked codegen, seems that no special handling is needed.