RFR: Avoid stub when calling to write-barrier from C2, remove all related code

[Roman Kennke]

I'd like to propose to let C2 call directly to write-barrier runtime,
instead of going via the stub. The original reason for going through the
stub was to do the cset-check there (which we now do in C2 IR), and to
be able to use XMM spills are WB calls in fast path. We used to see
fairly bad regressions in serial benchmark without that.

Given that:
- We're basically working around limitation in C2 RA
- We're adding a *lot* of code, some of it in shared paths (that we'd
need to upstream somehow)
- I don't see much impact when simply doing the right thing (see below)

I would like to propose to revert and get rid of all that fluff and do
the straight call using CallLeaf to ShenandoahRuntime::write_barrier_JRT().

I checked the serial benchmark and it seems to be consistently 1% worse
with this patch. Other benchmarks (specjbb/specjvm) seem unaffected.

I'd encourage you (esp. Aleksey) to run some benchmarks of your own to
check this. I don't want to push anything that causes regressions.

http://cr.openjdk.java.net/~rkennke/remove-c2-wb-stub/webrev.00/

Testing: tier3_gc_shenandoah, selected benchmarks

Thoughts?

Roman