Add addConsLinear() to avoid creating expressions in MILPs

[Joao-Dionisio]

I've been running some experiments comparing the performance of different Python interfaces. PySCIPOpt lands in a pretty nice second place, behind only OR-Tools. Still, why not try to get first place? :)

For each nonzero, quicksum(a * x[j] for ...) builds a new Expr, hashes a Term, and inserts it into a dict. addCons then goes through ExprCons and iterates that dict again to fill the C arrays. OR-Tools does a single C++ call per coefficient (ct.SetCoefficient(x[j], a)) and creates no Python objects.

In this PR, I'm adding an addConsLinear() function, similar to what addConsKnapsack() and addConsSOS1() do, adding the constraints directly while skipping the expression building.

I locally timed the building 13 MIPLIB 2017 instances with more than 100k nonzeros each. Here are the results:

instance	vars	conss	nnz	quicksum + addCons (s)	addConsLinear (s)	OR-Tools (s)
geometric mean				0.706	0.181	0.312
30n20b8	18,380	576	109,706	0.048	0.022	0.024
neos-1122047	5,100	57,791	163,640	0.193	0.046	0.060
atlanta-ip	48,738	21,732	257,532	0.173	0.080	0.081
rail01	117,527	46,843	392,086	0.355	0.181	0.151
supportcase6	130,052	771	584,976	0.351	0.183	0.177
roi2alpha3n4	6,816	1,251	878,812	0.249	0.032	0.132
neos-5049753-cuanza	242,736	322,248	1,440,672	1.50	0.517	0.525
ns1644855	30,200	40,698	2,110,696	0.727	0.112	0.360
supportcase7	138,844	6,532	2,845,545	0.998	0.265	0.503
neos-5052403-cygnet	32,868	38,268	4,898,304	1.52	0.191	0.783
neos-2075418-temuka	122,304	349,602	7,610,261	3.24	0.689	1.42
neos-3402454-bohle	2,904	2,897,380	8,953,800	10.22	2.45	3.01
square41	62,234	40,160	13,566,426	3.99	0.538	2.05

PS: In my tests, the matrix API is not faster than just using normal loops.

@Zeroto521 would you mind taking a look, please?

[Joao-Dionisio]

Despite the numbers we started from being already pretty small, this often matters in industrial applications. We academics may be used to the actual solving to take most of the solving time, but often the problems are actually super easy if presolving detects a certain structure, they're just huge.