Add support for newline-delimited JSON

[eyeseast]

☝️ does that. Closes #12

A couple caveats:

• Features are streamed in with a generator, since this is intended for large datasets. That means we can't look ahead at the first 100 features.
• We can't auto-detect feature IDs using ndjson. I'm ok with that. You can still pass --pk=id and get the same thing.
• We can't use the first 100 features to build the initial table. Again, I think it's fine. I note in the README how to grab a subset using Fiona.

I have one test here, but I'm very much open to suggestions for more.

[simonw]

There's a pattern for peeking ahead in sqlite-utils here:

https://github.com/simonw/sqlite-utils/blob/e8b2b7383bd94659d3b7a857a1414328bc48bc19/sqlite_utils/db.py#L993-L1004

You can use itertools.islice to pull out the first 100 items (and turn them into a list with list()), then use itertools.chain(that_list, original_iterator) to loop through the first 100 items followed by the rest of the iterator.

[simonw]

This looks great - the tests look robust enough to me.

[eyeseast]

Cool. I'll see if I can get the peek-ahead to work in a reasonable way. Was thinking about islice but wasn't totally sure where to do that with yield_records also happening.

[eyeseast]

I got feature.id working by sampling the stream of features coming in, and then chaining that sample back into the original stream.

Getting it to work with processed features to guess column types starts to feel precarious, because lists and generators are going to operate differently. It might ultimately be easier to do features = iter(features) at the top, so it's always a one-way stream and everything operates the same, but I'm not sure that's worth it. I think collecting a subset into a feature collection, like I describe in the readme, actually feels a little easier and more deliberate.