PERF: remove use of Python sets for interpolate

[simonjayhawkins]

have been investigating avoiding internals for interpolation. xref #34628

This PR address the issue of using Python sets, which is responsible for the bulk of the time in our current asv.

There are other improvements (will raise other PRs), but would need new benchmarks to show a benefit such as different index types and unsorted indexes. so this PR is targeting our current benchmark first.

prelim results (will post asv results if tests pass)

N = 10000
# this is the worst case, where every column has NaNs.
df = pd.DataFrame(np.random.randn(N, 100))
df.values[::2] = np.nan
df
%timeit df.interpolate()
# 189 ms ± 24.1 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)  <-- master
# 65.1 ms ± 635 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)  <-- PR

[simonjayhawkins]

some environments are failing with TypeError: diff() got an unexpected keyword argument 'prepend'

[simonjayhawkins]

we could move the above logic into _interp_limit and fastpath if limit_area.

although the algorithm could probably be extended to 2d, in which case the limit_area logic below would also need to move.

[simonjayhawkins]

although the algorithm could probably be extended to 2d, in which case the limit_area logic below would also need to move.

i don't think a 2d version will be needed as interpolate_2d will also be applied along axis when max_gap is added.

we could move the above logic into _interp_limit and fastpath if limit_area.

probably best as a follow-on to keep the diff here smaller

[simonjayhawkins]

will move this into a top level function called ffill_mask_with_limit (in this PR)

and plan to take the max_gap logic from #25141 and have a analogous ffill_mask_with_max_gap (separate PR)

[jreback]

looks fine.

[jreback]

do we have sufficient asv's for this?

[simonjayhawkins]

will get back to this after 1.0.5 we don't have many asvs so could probably add more The current asv is a single float block with a sorted RangeIndex and no limit (which can be fastpathed) so not representative of potentially much slower cases

[simonjayhawkins]

I'll close this for now. still don't have access to desktop and running asvs on laptop is painful.