attempt to share loader for JIT-ed classes

[kares]

introduces a -Xjit.loader.mode which is either UNIQUE (default), SHARED or SHARED_SOURCE

most of it is making WeakValueMap implement Map as I originally did something else here ...
so I've kept that code and just did a fake Map for the single (shared) global class-loader.
that can be changed as needed to a cleaner (proper oop) way.

the SHARED_SOURCE was so far only tested with jruby's test suite, the JITed class-loader count went down to 10% ... and meta-space was much less fragmented - saving around 8% of allocated space.

was not sure whether we should attempt to somehow detect eval-ed pieces
and not share them with the same source file but rather always on their own?
given that SHARED_SOURCE isn't changed to be the new default, we can also decide later.

also if you have suggestions for better naming the option or the enum values.

[headius]

This all seems good to me!

I don't know what to say about eval methods; in order for them to be a problem, they'd have to:

• Be continually generated at runtime. This is very uncommon.
• Be called enough to JIT. This seems unlikely now and even more unlikely with our time-based threshold changes.

On the flip side, many methods get generated using eval in a "safe" way (only once, usually to lazily metaprogram some methods) and we definitely want those to be eligible for JIT without a whole separate classloader (when running in a shared mode).

I think we go with this for now as-is. Folks that opt into shared classloading can help us uncover any weird eval edge cases.