[GlobalOptimization] Make im2col a generic convolution rewrite (#25011)

Replaces the three layout-specific im2col patterns (NHWC, NCHW,
depthwise NHWC) with one generic rewrite that handles any convolution,
named or generic, including grouped, dilated, 1-D/3-D, dynamic and
integer ones.

The pattern gathers the convolved input with an all-parallel
`linalg.generic`, rewires the op to read the gather, keeps the original
payload, and collapses the result into a matmul using
`DispatchCreation::getCollapsibleLoops`. Add a basic heuristic via a
control function that skips convolutions without an output channel
dimension (depthwise, pooling), where the gather is never amortized.

The pass is renamed to `iree-global-opt-convert-conv-to-im2col`; the
pipeline flag is unchanged.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Signed-off-by: Thomas Ziereis <ziereis@roofline.ai>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>