[TKW] Hoist loop invariant reads

[raikonenfnu]

In flash attention, Q's reduction dimension is typically relatively small, and hence we do only have reduction tile across K2 dimension/reduction dimension of 2nd gemm(P and V). Hence, an optimization we can do is to hoist reading of Q from global memory out of the for loop, this actually generates quite a big speedup (hoistQ + use global->register for Q gives typically 2x speed up.)

To implement the optimization above, we needed to add:

1. Expand hoisting.py to also look for Read that is:
   • independent of induction variable
   • has memory that is read-only/do not have write as users (i.e important for correctness since this guarantee data being read is constant/not changing with loop)
   • has memory who is a captured_var (i.e memory can be traced to outside the loop)
2. Implement method to hoist reads properly:
   • Copy Read to rootOp
   • replace rootOp's memory who is a captured_var with it's counterpart in the RootOp by querying reduction's implicit_capture
   • Remove unused captured_var from Reduction otherwise scf.for will be indexing/loading from the wrong bindings.
3. Updated specifically chained_gemm_tests in lit_tests/codegen.py to test for the hoisted reads from global.
4. Updated lit_tests/attention.py since this change generates new schedule