Fix: MPS reader bound conventions, infinite bounds, and unsupported sections

[ZedongPeng]

Description

The MPS reader deviated from the standard MPS conventions in several ways, each of which silently produced a different LP than other solvers read from the same file. Two numerical checks that other solvers perform were also missing.

1. MARKER 'INTORG'/'INTEND' was ignored. By the classic MPS convention (CPLEX, Gurobi, SCIP, HiGHS), an integer column with no BOUNDS entry is [0, 1]; we gave it [0, +inf). Any BOUNDS entry cancels the default. LI/UI, previously skipped, are now accepted.
2. A negative UP bound did not release the lower bound, giving the empty interval [0, -3] for UP x -3. It now implies lb = -inf, but only when no explicit lower bound was given.
3. A large finite bound was taken literally, polluting rescaling and the termination criteria. Bounds at or beyond +/-infinite_bound (new parameter, 1e20 by default, as in HiGHS and Gurobi) are now replaced by true infinities. This lives in preprocess_problem rather than the reader, as in HiGHS, so models built through the Python or C API get it too.
4. A huge matrix or objective coefficient went unreported. Either wrecks the scaling and is almost always a modelling mistake. preprocess_problem now warns about matrix entries reaching 1e15 and objective coefficients reaching 1e20 (HiGHS's thresholds). HiGHS errors instead; we only warn, having no "invalid input" termination status to return.
5. A file with no N row stole the first constraint as the objective, and sections that change the model beyond an LP (QUADOBJ, CSECTION, INDICATORS, ...) fell through into whichever section was active. A missing N row now gives a zero objective, those sections are rejected, and SOS/SETS are skipped since they leave the LP relaxation unchanged.

Effect

Only (1) changes any instance we solve - 12 in MIPLIB 2017:

ds, ds-big, eilA101-2, eilC76-2, ivu06, ivu06-big, ivu52, ivu59, neos-1354092, neos-3352863-ancoa, ns2122698, rvb-sub

Everything else in our benchmark set parses byte-for-byte identically. The 12 were re-solved at 1e-4 and 1e-8 and still terminate OPTIMAL with essentially unchanged objectives: these are set-covering-style models whose constraints already imply x <= 1, so the restored bound is redundant.