algorithms: reduce memory footprint of local UDFs

Created by: jassak

Some memory expensive lines are identified, in local UDFs, and are replaced with equivalent memory efficient ones.

All changes fall into three categories:

• Avoid memory copies when operation can be done inplace
• Use of numpy.einsum, when applicable, for summing over arrays
• Avoid using sklearn methods when re-implementing is easy, because sklearn hasn't been written with the above two considerations in mind

There is also one change in udfio.py. When constructing a pandas.DataFrame from separate columns, for relational tables, the flag copy=False is now used. This avoids consolidating columns into a single 2-dimensional array. However, when some operation, later in the UDF, requires the columns to be consolidated, pandas will eventually do it in the background, copying their data into a single contiguous block of memory. This is unavoidable since many operations in numpy need to operate on contiguous blocks of memory in order to maximize their efficiency, since they exploit the CPU cache and the CPU vectorization capabilities.

In most cases the memory reduction is accompanied by a reduction in execution time as well.

Results are presented in the next table

file	method	data	mem before	mem after	perc	time before	time after	perc
pca.py	local1	160MB	172MB	2KB	0.001%	250ms	30ms	12%
pca.py	local2	160MB	305MB	153MB	50%	300ms	400ms	133%
pearson.py	local1	320MB	153MB	5KB	0.003%	330ms	300ms	90%
ttest_independent.py	local_independet	16MB	24MB	3KB	0.012%	360ms	190ms	53%
ttest_onesample.py	local_one_sample	8MB	16MB	3KB	0.018%	380ms	800μs	0.21%
ttest_paired.py	local_paired	16MB	24MB	8MB	33%	600ms	4ms	0.67%
metrics.py	_confusion_matrix_local	16MB	44MB	2MB	4.5%	140ms	6ms	4.3%
metrics.py	_roc_curve_local	16MB	1.6GB	2MB	0.12%	2.7s	600ms	22%
logistic_regression.py	LogistcRegression._fit_local_step	24MB	24MB	8KB	0.32%	23ms	233ms	1000%
linear_regression.py	LinearRegression._compute_summary_local	16MB	8MB	8MB	100%	360ms	5ms	1.3%
anova_oneway.py	local1	16MB	106MB	74MB	70%	200ms	100ms	50%
destriptive_stats.py	local	168MB	424MB	358MB	84%	6s	5.8s	97%
udfio.py	from_relational_table	2GB	2GB	8KB	~0%	1.6s	150μs	0.009%