PERF: faster unstacking

[jreback]

closes #15503

so on a non-masked unstack (IOW, a fully product multi-index for example), this is now just
a simple reshape. On a masked unstack, it now will have a much lower O constant, as its in cython, and with release the GIL.

0.19.2 / master

In [2]: m = 100
   ...: n = 1000
   ...: 
   ...: levels = np.arange(m)
   ...: index = pd.MultiIndex.from_product([levels]*2)
   ...: columns = np.arange(n)
   ...: values = np.arange(m*m*n).reshape(m*m, n)
   ...: df = pd.DataFrame(values, index, columns)
   ...: 

In [3]: %timeit df.unstack()
1 loop, best of 3: 285 ms per loop

In [4]: df2 = df.iloc[:-1]

In [5]: %timeit df2.unstack()
1 loop, best of 3: 306 ms per loop

PR

In [2]: %timeit df.unstack()
10 loops, best of 3: 70 ms per loop

In [3]: df2 = df.iloc[:-1]

# & releasing the GIL here.
In [4]: %timeit df2.unstack()
1 loop, best of 3: 191 ms per loop

[codecov-io]

Codecov Report

Merging #15510 into master will decrease coverage by -0.03%.
The diff coverage is 100%.

@@            Coverage Diff             @@
##           master   #15510      +/-   ##
==========================================
- Coverage   91.04%   91.02%   -0.03%     
==========================================
  Files         136      136              
  Lines       49088    49105      +17     
==========================================
+ Hits        44694    44698       +4     
- Misses       4394     4407      +13

Impacted Files	Coverage Δ
pandas/core/reshape.py	99.27% <100%> (+0.02%)	✅
pandas/io/gbq.py	25% <0%> (-58.34%)	❌
pandas/tools/merge.py	91.78% <0%> (-0.35%)	❌
pandas/util/testing.py	81.87% <0%> (-0.19%)	❌
pandas/core/frame.py	97.82% <0%> (-0.1%)	❌

---

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 cd67704...ec29226. Read the comment docs.

[jreback]

cc @wesm if you have a chance.

[wesm]

Very nice. Don't love expansions of our sprawling Cython codebase, but this seems like a solid win as a pretty central data manipulation.

[jbrockmendel]

@jreback couldn use some help grokking how we get here. In Block._unstack adding an assertion assert mask.all() doesn't break any tests. is that something we can rely on? (if so we can simplify code a good bit) If not, how can we construct a counter-example?