test string casting only on first element

[lvwerra]

This addresses #153: Instead of testing on every element of a batch we only test on the first. This should reduce the latency for adding batches.

This makes the assumption that the user would pass the same type in the rest of the batch which is a fair one IMO.

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[lhoestq]

I think for #153 the matter is about checking the first non null element in a Sequence feature type, and not about checking the first element of a batch (though it can be complementary)

This can be done inside _enforce_nested_string_type

[lvwerra]

True, I guess it depends what is larger: samples ( $n$ ) in dataset or sequence length ( $s$ ). I also added that only the first element of a sequence is tested. So with this change we went from $O(n \times s)$ to $O(1)$. This decreased the time of the example in #153 from 24 to 6 seconds on my machine. You can also see in the profile that _enforce_nested_string_type is now a tiny fraction (~0.1%) of the computation (it's the narrow pink bar on the right).

cc @NouamaneTazi

[NouamaneTazi]

LGTM. I think we could clarify in the docs that using Array..D is still faster if we're working with fixed length arrays.

[lhoestq]

LGTM :)

(feel free to write some tests btw)