Feature: optional forward diff mode

[mattia-spider]

Add optional forward-mode autodiff to diff() for cases where dim(input) << dim(output).

Selectable via mode="fwd" (default "rev", so existing uses like force computation are unchanged). Forward mode drastically reduces memory overhead when differentiating w.r.t. a low-dimensional input (e.g. the 3-component external field for NMR tensors), since it needs O(n_input) passes instead of O(n_output).

Memory impact, same architecture on qm9-nmr:

80GB H100 (with batch size 8): peak_MiB: 35477 (rev) → peak_MiB: 4849 (fwd)

8GB GPU: max batch 3 (rev) → 11 (fwd)
80GB H100: max batch 8 (rev) → 40 (fwd)

Correctness: fwd and rev give the same Jacobian, max abs diff ~3e-6 (float32), ~1e-15 (float64).
TODO (lines 297-303): forward mode breaks e3nn IrrepsArray's shape invariant (jacfwd appends the tangent axis, colliding with the irreps dim on rewrap). Worked around with value = base.as_array(value), active only in fwd mode. Proper fix belongs upstream in e3nn; remove workaround once resolved.

[codecov-commenter]

⚠️ Please install the to ensure uploads and comments are reliably processed by Codecov.

Codecov Report

❌ Patch coverage is 87.50000% with 1 line in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/tensorial/gcnn/_diff.py	87.50%	1 Missing ⚠️
❗ Your organization needs to install the Codecov GitHub app to enable full functionality.

Files with missing lines	Coverage Δ
src/tensorial/gcnn/_diff.py	79.55% <87.50%> (ø)

🚀 New features to boost your workflow:

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

[muhrin]

Great! Could you try removing your x.array workaround for forward mode together with my e3nn-jax version: https://github.com/muhrin/e3nn-jax/tree/develop

I think I've fixed the bug affecting the Jacobian.