Use microbatch B>1 not possible

Tracking issue for the work needed to enable microbatch dimension `B > 1` in ODT.

Today the framework operates with implicit `B = 1`: each microbatch is a single sample, and shapes like `[1, F]` are common. Several places encode this assumption explicitly or implicitly:

- Loss-API uniform-microbatch contract (`docs/CONVENTIONS.md` → "Loss API: microbatch contracts" — landed with #135)
- DataLoader returns one sample per batch index in `getBatch` — shape `[1, ...]` not `[B, ...]`
- Layer forward/backward implementations may not support batched input
- Training-loop per-microbatch counters assume B = 1

When `B > 1` becomes a real feature target, the sub-issues below capture the verified-needed work. New sub-issues land here as we discover them — this is the umbrella, not the implementation bag.

## Sub-issues

- [ ] #153 — Microbatch shape contract asserts in loss forward/backward

## Out of scope

This issue does not get implemented as a single PR. It is a tracking umbrella; each sub-issue lands independently.