PERF: Improve efficiency of BlockValuesRefs

[Tolker-KU]

• closes #xxxx (Replace xxxx with the GitHub issue number)
• Tests added and passed if fixing a bug or adding a new feature
• All code checks passed.
• Added type annotations to new arguments/methods/functions.
• Added an entry in the latest doc/source/whatsnew/vX.X.X.rst file if fixing a bug or adding a new feature.

Improve efficiency of BlockValuesRefs by allowing for more efficient C code-gen

[rhshadrach]

@Tolker-KU - can you run and post ASV results

[Tolker-KU]

@Tolker-KU - can you run and post ASV results

Thanks for taking a look a this.

I've tried hard to get meaningful runs of the asv benchmarks but I think my computer is so old and slow that noise is dominating the results. I'm getting weird results where some benchmarks are much faster and some are much slower

To show the gains from this I've run below micro-benchmark

>>> N = 10_000
>>> arr = np.random.normal(size=(50, N))
>>> df = pd.DataFrame(arr)
>>> %timeit -n10000 list(df.iterrows())
2.43 ms ± 268 μs per loop (mean ± std. dev. of 7 runs, 10,000 loops each)  <- before
2.1 ms  ± 117 μs per loop (mean ± std. dev. of 7 runs, 10,000 loops each)  <- after

[WillAyd]

Does this actually change anything? I am under the impression that Cython would be smart enough to handle this the same in both cases.

Instead of looking at the benchmarks (which can be difficult to run and flaky), you can also inspect the Cython annotations before/after the change. Are they actually different?

[Tolker-KU]

Thanks. Glad you brought that up. I should've been more explicit in the description, as the motivation for this PR is that the code-gen changes quite dramatically. An example from one of the methods below. Note the code generated for the method does not fit on one screen before the change

Before

After

[WillAyd]

Thanks for posting. That's...quite the difference

[mroeschke]

Thanks @Tolker-KU