test(backward): assert a batched invoke's gradient is its own rows

[Hotragn]

The batched-gradient half of #699, on its own, per your review there:

If you want to send the batched-gradient coverage on its own, the parametrisation over which invoke takes the gradient and the differing row counts so a wrong offset fails on shape, that stands by itself and I would take it. Nothing currently asserts that a batched invoke's gradient is its own rows.

No diagnostic, no docs — those are settled. Tests only, no source changes.

What it pins

Invokes share one batched forward, so a module's output is a single tensor holding every invoke's rows. Reading .grad inside an invoke has to narrow to that invoke's slice; reading the whole batch, or another invoke's slice, gives a plausible-looking tensor that is silently the wrong gradient.

Four cases, all against reference_grad — what plain autograd says for that input alone:

• Parametrised over which invoke takes the gradient, first through last of three. Each sits at a different batch offset, so only exercising all of them catches a slice taken from the wrong place.
• Row counts differ per invoke (2 / 3 / 1), so a wrong offset fails on shape rather than on values that might round close.
• A same-row-count pair, where the shape check cannot be what catches a wrong offset — only the values can. Asserts the gradient read is not the other invoke's.
• Two invokes each taking a gradient, and an edit in one leaving the other's rows untouched.

Confirming they have teeth

Since nothing asserted this before, I checked against a mutant that forces the gradient to be read from batch row 0:

-geometry = (tensor.shape, tensor.stride(), tensor.storage_offset())
+geometry = (tensor.shape, tensor.stride(), 0)

5 failed, 8 passed

Every new test whose target is not row 0 fails — including the same-row-count pair, which catches it on values alone. All seven pre-existing tests in the file still pass under the mutant, which is the gap this fills.

One judgement call

_BatchEnvoy is duplicated from test_batching.py rather than imported across test modules — that file is deliberately all @torch.no_grad(), and these are the gradient tests. Happy to hoist it to conftest.py instead if you would rather have one copy; I left it local rather than touching a shared file unasked.

Verification

tests/test_backward.py: 13 passed (7 existing + 6 new).

_{0.8 @ fe8645c, Python 3.14.3, transformers 5.16.1, torch 2.9.1, CPU.}