PERF: combine ops can be block based

[jreback]

related #5482 / #5481

in theory an op like:

df * df2 can be done block-based rather than column-by-column, so for large numbers of columns would see nice speedups. Though, passing to pd.eval is almost always faster because its passed to numexpr (at the cost of an upcasted dtype on the way back)

the 2 frames are aligned so that they have the same columns; then need to cast
each of the other (df2) in this case to a joint common type (happens in any event when you operate.

Then can just do block-by-block ops

prob BlockManager.eval can handle this with a kw arg to distinguish

[jbrockmendel]

Can this be closed as Sufficiently Stale? If not, -1 on moving anything in to internals.

[jbrockmendel]

@jreback I think this can be closed. We’ve now gotten rid of Block.eval entirely. I’ll be putting some effort into optimization in the near future, but not via BLock dispatch.