computeSliceParameters: use full slice if affine exprs are non-monotonic

[mgehre-amd]

computeSliceParameters maps the loop range (given by lower bounds lbs and upper bounds ubs) into the operand tensor to compute the input slices that contains all indexed values.

The computation currently assumes that the affine expressions to map from loop iteration domain to tensor domain are monotonic, and use that information by just evaluating those at lbs and ubs to obtain the slice o the operand tensor.
For non-monotonic expressions, the maximum and minimum values might not be at the boundaries.
Detect that case, and then use the full slice to be safe.

Fixes llvm#111830

If shapes are static, we could be cleverer: sometimes we can prove that even though the affine expression is not monotonic in itself, it is monotonic on the range (lbs, ubs). For example when the affine expression contains a d0 mod 8, and the range is (0, 4).

[cferry-AMD]

LGTM, thanks for solving! That's much simpler than using the LexMin and LexMax which I imagined as a fix for that.