feat: add fold4d / Fold4d module (inverse of unfold4d)

[Bellman281]

Pull Request Template

Checklist

• Confirmed that cargo run-checks command has been executed.
• Made sure the book is up to date with changes in this PR.

Related Issues/PRs

Provide links to relevant issues and dependent PRs.

Changes

Summarize the problem being addressed and your solution.

Testing

Describe how these changes have been tested.

[Bellman281]

Checklist

• Confirmed that cargo run-checks command has been executed.
• Made sure the book is up to date with changes in this PR.

Related Issues/PRs

Closes #786.

Changes

Adds a fold operation (col2im), the inverse of the existing Unfold4d / unfold4d.

• burn_tensor::module::fold4d(x, output_size, kernel_size, options) — combines an
  array of sliding local blocks [N, C * kH * kW, L] back into a tensor
  [N, C, H, W], summing the values of blocks that overlap. Matches
  torch.nn.functional.fold.
• Fold4d / Fold4dConfig in burn-nn, mirroring Unfold4d, exposing
  output_size, kernel_size, stride, padding, and dilation.

Implementation. fold4d is the adjoint of unfold4d. Since unfold4d is a
conv2d with a one-hot "im2col" kernel, fold4d reuses that exact kernel through
conv_transpose2d, with padding_out = (size + 2·pad − dil·(k−1) − 1) mod stride to
recover the requested output_size. It is composed entirely from existing ops, so the
backward pass comes for free through autodiff and no backend / IR / router changes are
needed.

No book change: Unfold4d has no entry in the module reference, so there is nothing to
mirror for Fold4d.

Testing

• Known-answer test: folds a [1, 4, 4] column tensor into a 3×3 output with a
  2×2 kernel; expected values were computed independently with a numpy col2im
  reference.
• Round-trip test: verifies fold4d(unfold4d(x)) equals a numpy-computed reference
  (overlaps summed), confirming fold is the true adjoint of unfold.
• The conv_transpose2d + padding_out approach was cross-checked against a direct
  col2im implementation across 7 configurations (varying stride, padding, dilation,
  and channels), all matching exactly.
• cargo run-checks passes (fmt, clippy, typos, tests).

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 66.03%. Comparing base (bd6e8fa) to head (8705717).
⚠️ Report is 6 commits behind head on main.

❌ Your project check has failed because the head coverage (66.03%) is below the target coverage (80.00%). You can increase the head coverage or adjust the target coverage.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #5173      +/-   ##
==========================================
+ Coverage   65.97%   66.03%   +0.06%     
==========================================
  Files        1198     1203       +5     
  Lines      182691   183118     +427     
==========================================
+ Hits       120526   120921     +395     
- Misses      62165    62197      +32     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[Bellman281]

Thanks! All CI checks are green (macOS just finishing up), and codecov/patch is 100%.
@Charles23R @nathanielsimard whenever you have a moment, this is ready for review — it adds fold4d / Fold4d as the inverse of Unfold4d (closes #786), built as its adjoint via conv_transpose2d so autodiff is free and no backend changes were needed. 🙏