It appears there's significant overhead in directly obtaining the name
of every sub-element using a loop of `h5?_get_name_by_idx` calls.
For instance, a v7.3-formatted Matlab save file I have access to has
deeply nested and large struct which are serialized into a huge number
of datasets within a `#refs#` group.

```julia
julia> Int(length(fid["#refs#"]))
36236
```

Before this change, enumerating the keys takes almost a minute:
```julia
julia> @time keys(fid["#refs#"]);
 54.193914 seconds (72.47 k allocations: 3.594 MiB)
```
but with this change, that's reduced to a fraction of a second:
```julia
julia> @time keys(fid["#refs#"]);
  0.013353 seconds (36.24 k allocations: 1.383 MiB)
```