Check for permuted indices in NL segments

[vitaut]

Ill-formed .nl files can have indices permuted. This doesn't affect the NL reader because it passes the indices to the client, but NLProblemBuilder should detect it and do something about it like reporting an error.

[vitaut]

In the C, O and F segments the order doesn't matter and duplicates will overwrite older entries which is OK.

[vitaut]

Segment that need checks: V (common expressions) and L (logical constraints).

[vitaut]

Done for logical constraints (L).

[vitaut]

Looks like adding all objectives at once and then updating them with expressions and types is faster (and easier) than adding them incrementally and checking if they are ordered:

Benchmark          Time(ns)    CPU(ns) Iterations
-------------------------------------------------
AddObjsAtOnce            39         39   22790390                                 
AddObjsIncVector        170        174    3537068                                 
AddObjsIncArray          85         88    7982500 

https://github.com/ampl/mp/blob/master/test/nl-benchmark.cc

[vitaut]

Done for objectives.

[vitaut]

Done for constraints. Need to do the same for common expressions and functions.

[vitaut]

Done for common expressions.

[vitaut]

Also check if avoiding push_back indeed improves performance as the benchmark suggests.

[vitaut]

All permuted indices are handled correctly now.

[vitaut]

Remaining work here:

• Check if avoiding push_back improves performance
• Make sure that NLProblemBuilder is fully tested

[vitaut]

Here are nl-benchmark results on https://github.com/ampl/mp/tree/3cd123f67fb368cc3ecf49ddbd0c461ac3b750d7 (Dec 30, 2014):

Binary:

Method	Time, s
nl reader	0.24
nl reader+build	1.60
ASL	2.16

Text:

Method	Time, s
nl reader	1.31
nl reader+build	2.71
ASL	3.48

Tested on machine with the following specs (titan):
CPU: Intel(R) Core(TM) i7-5820K CPU @ 3.30GHz
RAM: 7885MiB

[vitaut]

Results on https://github.com/ampl/mp/tree/75791139a06eb8e021b6493fe46a863956f2bb8a (Mar 25, 2016):

Binary:

Method	Time, s
nl reader	0.24
nl reader+build	1.60
ASL	2.16

Text:

Method	Time, s
nl reader	1.26
nl reader+build	2.69
ASL	3.49

Text slightly improved, binary the same.

[vitaut]

Results on e6afdba with fixed handling of permuted indices:

Binary:

Method	Time, s
nl reader	0.23
nl reader+build	1.60
ASL	2.15

Text:

Method	Time, s
nl reader	1.27
nl reader+build	2.70
ASL	3.44

[vitaut]

Avoiding push_back doesn't improve performance for nonlinear problems because the time is probably dominated by other factors (expression allocation?), but at least there is no need to construct temporary arrays to permute indices.