[matthiaskoenig]

LLJit reused the object of a model compiled earlier in the process
whenever the md5 of the SBML matched. The key did not hold the load
options, so a model loaded with another LLJIT_OPTIMIZATION_LEVEL kept
the code of the first level, and a model loaded with other options,
e.g. MUTABLE_INITIAL_CONDITIONS, got code without the functions it
needs and crashed. LOADSBMLOPTIONS_RECOMPILE did not prevent the reuse.

getSBMLMD5 now returns a key of the md5 of the SBML, the load options
except RECOMPILE (they include the JIT and its optimization level) and
LOADSBMLOPTIONS_PERMISSIVE, which the code generation reads. The model
cache of the LLVMModelGenerator and the object cache of LLJit use it,
and the copy of getSBMLMD5 in the LLVMModelGenerator is gone. Each LLJit
has its own object cache in front of the shared one: with RECOMPILE it
compiles the module instead of looking for it, and it keeps the object
of its model, which saveState uses instead of looking up the shared
cache, where another model may have replaced it. The shared cache holds
shared objects and locks its lookups too.

Fixes #1420.

With develop, this crashes in the second RoadRunner with "Could not find symbol getFloatingSpeciesAmount", and works with this change:

import roadrunner
C = roadrunner.Config
C.setValue(C.LLVM_BACKEND, C.LLJIT)
C.setValue(C.LOADSBMLOPTIONS_MUTABLE_INITIAL_CONDITIONS, True)
roadrunner.RoadRunner("feedback.xml")
C.setValue(C.LOADSBMLOPTIONS_MUTABLE_INITIAL_CONDITIONS, False)
roadrunner.RoadRunner("feedback.xml")

The script of #1420 now compiles the model three times, once for each load.

[luciansmith]

Recreated in #1440.