Deprecate 'with dset.astype()'

[takluyver]

h5py 3.0 introduced the ability to slice an astype wrapper directly (ds.astype(np.int32)[:]). I'd like to remove the old pattern of using astype() as a context manager with mutable state, and get rid of the thread-local storage it requires.

Edit: I am happy if we want to give people more time to update to 3.x before adding this warning. We released 3.0 in October 2020. The 3.3 release may be around 6 months after that, depending on how eager we are to do it.

[codecov]

Codecov Report

Merging #1842 (c1bdb52) into master (1340914) will increase coverage by 0.01%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master    #1842      +/-   ##
==========================================
+ Coverage   88.38%   88.39%   +0.01%     
==========================================
  Files          17       17              
  Lines        2281     2284       +3     
==========================================
+ Hits         2016     2019       +3     
  Misses        265      265              

Impacted Files	Coverage Δ
h5py/_hl/dataset.py	92.56% <100.00%> (+0.03%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 1340914...c1bdb52. Read the comment docs.

[takluyver]

Add the warning for 3.3, or wait a while longer?

This is trickier than #1839. The sooner we start pushing people away from this, the sooner we can get rid of the thread-locals piece. But the only way to keep code compatible with h5py 2.x and 3.x is to use the older style (with dset.astype(np.int32):). This also means it's necessary for any code that still has to support Python 2.

I'm leaning towards leaving this a bit longer - maybe add the warning a year after h5py 3.0 was released.

Changed my mind about timeline.

[tacaswell]

I'm sold that we should let people have a 12 month support window for older versions of h5py. This should wait until Nov 2021 (which I presume will be h5py3.4).

[aragilar]

Should we merge this now?

[takluyver]

Yes, it has been a year since 3.0 was released with the new arr = dset.astype(np.int32)[:10] pattern, which was the delay we agreed on before warning. We can allow it with a warning for another 6 or 12 months, I guess.