Tracking: untracked TODOs and broken tests in src/ and tests

This is a tracking issue consolidating in-code TODOs and `@test_broken` tests that currently have no corresponding GitHub issue. Found via an audit of `src/` and `test/`.

## TODOs in `src/`

- [x] **`chunk` rrule captures `x` in the pullback** — [`src/utils.jl:219`](https://github.com/JuliaML/MLUtils.jl/blob/main/src/utils.jl#L219): `# TODO avoid capturing x in the pullback`
- [x] **`randobs` cannot take an RNG** — [`src/randobs.jl:1`](https://github.com/JuliaML/MLUtils.jl/blob/main/src/randobs.jl#L1): `# TODO: allow passing a rng as first parameter`
- [x] **No generic `show` for `AbstractDataContainer`** — [`src/datacontainer.jl:17`](https://github.com/JuliaML/MLUtils.jl/blob/main/src/datacontainer.jl#L17): `# TODO: add generic show method. See DataLoader's show for inspiration.`

## Broken tests in `test/`

### A. `eltype(DataLoader)` does not match the actual batch element type
`eltype(d)` disagrees with `eltype(collect(d))` across plain arrays, tuples, NamedTuples, and Dicts. 7 occurrences in [`test/dataloader.jl`](https://github.com/JuliaML/MLUtils.jl/blob/main/test/dataloader.jl):

- [x] lines 9, 20, 28, 37, 54 (Array / Tuple / NamedTuple cases)
- [x] lines 118, 126 (`Dict` cases, inside the `"Dict"` testset)

### B. `getobs` on a `Dict` is not type-inferred
- [ ] [`test/observation.jl:41`](https://github.com/JuliaML/MLUtils.jl/blob/main/test/observation.jl#L41): `@test_broken @inferred getobs(dataset, 2) # not inferred`

---

Each checkbox is an independent fix; this issue exists only to track them so the markers in the code are no longer orphaned.