[Dispatch] Make dynamic attention ineligible for collapse

[IanWood1]

For the same reasons we can't collapse dynamic linalg.generic ops, we also cannot collapse attention ops with dynamic dimensions. This moves the check for dynamic dims before checking for LinalgExt::AttentionOp. Saw causing collapse/expand ops in dispatches in #19921

[MaheshRavishankar]

For the same reasons we can't collapse dynamic linalg.generic ops, we also cannot collapse attention ops with dynamic dimensions. This moves the check for dynamic dims before checking for LinalgExt::AttentionOp. Saw causing collapse/expand ops in dispatches in #19921

why cant we collapse dynamic dimensions in attention op (cause I cant immediately remember why we cant collapse dynamic dimensions of linalg.generic. The only reason I can think of was that tensor.expand_shape didnt support dynamic dimensions, but now it does).

[IanWood1]

For the same reasons we can't collapse dynamic linalg.generic ops, we also cannot collapse attention ops with dynamic dimensions. This moves the check for dynamic dims before checking for LinalgExt::AttentionOp. Saw causing collapse/expand ops in dispatches in #19921

why cant we collapse dynamic dimensions in attention op (cause I cant immediately remember why we cant collapse dynamic dimensions of linalg.generic. The only reason I can think of was that tensor.expand_shape didnt support dynamic dimensions, but now it does).

There are some changes to the pass needed (see #19654). But in this case this might be overkill since the actual collapsed tensor is fully static even though the attention op has dynamic dims. The attention pattern just bails when any of the loops are dynamic.

[MaheshRavishankar]

How this work for LLaMa then? I thought that we previously had issues with attention op not being collapsed back down. If this doesnt cause regression on LLaMa (which isnt in CI), this looks fine to me.

[IanWood1]

Not needed after #19654