[LinalgExt] Implement direct vectorization for im2col op (#23855)

Implements direct vectorization for im2col, and completes the support
for padding on the im2col op. The padding attributes are used to compute
the read mask when vectorizing. In the old path, we would have separate
padding on the input and the result of the im2col, and we try to compose
those pads into a single masked read. This is fragile and difficult for
cases where the im2col result dims don't map well to the input dims.
With this direct vectorization approach, we can compute the mask based
on the input and result padding simultaneously. This will make
flattening of the spatial dimensions of convolutions possible.

### Performance results: ###

- Run 1:
https://github.com/nod-ai/amd-shark-ai-reports/tree/main/boo/boo-custom-runs/2026-03-27_04-36_d1b822f45ac693a8593232a7d3fc5d67b1087f7e/comparison
- Run 2:
https://github.com/nod-ai/amd-shark-ai-reports/tree/main/boo/boo-custom-runs/2026-03-27_20-34_cf0d758bb199713b96a84a67245bc8b24ba7b74a/comparison

These runs were taken on different commits, but they are functionally
the same (just some cleanup differences). I am only able to reproduce 3
of the regressions locally, and most of the improvers (~35 of them with
10-40% speedup) are real. There seems to have been some noise in the
runs, but overall there is a good perf improvement.

ci-extra: test_torch

---------

Signed-off-by: Max Dawkins <max.dawkins@gmail.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>