Fix type instability in unflattening Tuples

[simsurace]

On master this does not infer:

julia> tpl = (1., ((2., 3.), 4.));
julia> par, unflatten = flatten(tpl);
julia> @code_warntype unflatten(par)

This PR fixes this and (so I believe) improves the type inference with a recursive implementation and a fallback for empty tuples.

[willtebbutt]

Thanks for opening this. Busy morning -- will try to review later today

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Comparison is base (567b070) 97.36% compared to head (f5bdf67) 97.40%.
Report is 1 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master      #67      +/-   ##
==========================================
+ Coverage   97.36%   97.40%   +0.03%     
==========================================
  Files           8        8              
  Lines         228      231       +3     
==========================================
+ Hits          222      225       +3     
  Misses          6        6              

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[simsurace]

What is your view on using views (pun intended)?

[willtebbutt]

I think views definitely make sense here. Very open to that.

While you do that, I'm going to drop 1.5, because we really don't need it.

[simsurace]

Thinking it over a bit, I think going to views should be evaluated a bit more carefully, as it also requires generalizing a lot of the signatures of the unflatten functions. I will keep that for a later PR perhaps.

[simsurace]

I'm gonna suggest approving this as is. I bumped the version. Maybe we can release #68 in the same swoop?

[willtebbutt]

Thinking it over a bit, I think going to views should be evaluated a bit more carefully, as it also requires generalizing a lot of the signatures of the unflatten functions. I will keep that for a later PR perhaps.

Good point. Yeah, let's defer that until later. I do think it would be good to look at though, because we're currently doing a lot of copying.

[willtebbutt]

LGTM.

I'd really quite like to stick to the standard 1-version-per-PR rule though, so please tag this separately from #68

[simsurace]

No problem!