Make sure string column with vec size 1 work

[esheldon]

closes #426

This fixes a bug where the shape of the column was not correct.

Adds tests that string and number vec1 column data get returned, even though read as scalar

[erykoff]

Thanks!

[beckermr]

The code looks fine but I think I am missing some details on what the special case is exactly. I've asked for a few comments in places to help future us understand this edge case.

[beckermr]

I missed that this is a breaking change wrt the behavior of returning dims of (1,) versus a scalar. We should figure that bit out.

See also #426 (comment)

[esheldon]

It is a breaking change, in that those columns will now be properly round tripped whereas they were not before.

On the other hand, I consider this a bug in numpy that I couldn't get this to work in 1.X. Seen that way it is breaking in the sense of correcting behavior.

[beckermr]

Is there a way we can make it work for non-strings?

[esheldon]

There is no way to set the TDIM for length-1 vector columns, they must be length 2 or more.

I can see hacks working. We could add a header keyword that tells us the situation for a given column so fitsio can reshape the data.

[beckermr]

My feeling on this specific item is that this specific behavior of FITSIO is so ingrained in people's codes + minds that if we break it, we're going to cause silent or at least very weird bugs. I get that it can be considered a bug, but long-standing bugs that people code around effectively become part of the API.

[beckermr]

I can see hacks working. We could add a header keyword that tells us the situation for a given column so fitsio can reshape the data.

Right. My feeling is that we

1. Keep the current behavior before this PR where columns of shape (1,) are cast to scalars for all types.

2. If we do want to fix this, we fix it for all types AND we bump fitsio to major version 2.

This seems to me to be the only way to do this without causing lots of bad side effects.

[beckermr]

We could include some of the other big ideas we've discussed in version 2 as well, including the new I/O modes+features.

[esheldon]

I don't actually know how to workaround the new numpy behavior, will have to figure it out.

[beckermr]

Can't we just change the output dtype for strings if we find TDIM has 2 elements?

[esheldon]

There is code reuse for read and write. It seems like different behavior might be needed for read vs write in this case, but need to look into it.