RFR (M): 8011102: Clear AVX registers after return from JNI call

[Vladimir Kozlov]

http://cr.openjdk.java.net/~kvn/8011102/webrev/

A native library may use wide 256bit YMM registers and does not clean 
them after that causing SSE <-> AVX transition penalty.
Note: penalty exist only in mixing 256bit data avx instructions with SSE 
instructions. There is no penalty with mixed 128 bit avx instructions 
because they zero upper bits.

Add vzeroupper instruction after return from JNI call to avoid the
penalty.
Add vzeroupper in compiled code which use 256bit vectors, before calls 
and return. Also add it in intrinsics which use 256bit vectors: 
arraycopy, fillarray, strings and encode_iso.
Note, vzeroupper is fast, one cycle instruction.

Moved Java_To_Runtime() encoding to x86_64.ad from <os>_x86_64.ad 
because they are the same.

Tested with compiler jtreg, ctw, jprt.

Thanks,
Vladimir