[TKW] Modify Index Seq Analysis to handle "detours"

[raikonenfnu]

Our current index_seq_analysis, does a backward pass on lhs, rhs, and acc, and then does a forward pass on it's consumers. This is working out OK for now, however for more complex cases we may need to modify it to also do detours. Consider this case:

lhs = read
rhs = read
bias = read
mma = mma(lhs, rhs)
res = mma + bias
write(res)

In the case above, if we want read of bias to also have the layouts from mma's acc, we'd need to do a detour during layout setting i.e
mma -> res -(detour)-> bias read.

This is actually also evident in the case of our attention kernel. Currently, we are manually setting vector_shapes for M and N on our attention kernel

iree-turbine/tests/kernel/wave/wave_attention_test.py

Line 244 in 2b45c0f

vector_shapes={B: 0, M: 16, N: 16},

S.T the partial_sum/reduction of sum will have the correct expansion and indexing. In reality, we should be able to handle this by doing the detour.

[harsh-nod]

Does #292 address this issue?