PERF: put BlockManager constructor in cython

[jbrockmendel]

No idea why mypy started complaining about the io.parsers stuff

[jreback]

can you merge master

[jreback]

any perf on this vs master (now that the ndarray stuff is merged)? (ideally put the relevant timeits in the cython code, that is a great idea generally)

[jbrockmendel]

any perf on this vs master (now that the ndarray stuff is merged)? (ideally put the relevant timeits in the cython code, that is a great idea generally)

On the benchmark I've been using for these im seeing about a 4% gain (though the worst performance had a 0.2% slowdown). The big win I'm expecting comes in the next step, which puts BlockManager.get_slice in cython, which allows us to chain together several no-longer-python calls

[jbrockmendel]

rebased + green

[jbrockmendel]

Just measured the branch after this (that implements BlockManager.get_slice in cython) and that gets a 15% improvement.

[jorisvandenbossche]

On the benchmark I've been using for these im seeing about a 4% gain

As I have mentioned before (eg on the mailing list, in #40263), this benchmark is basically useless / unrealistic.
I think you should benchmark with an actual applied function to check wether it is worth moving things to cython for this.