[WIP] Remove unneeded JNI registrations.

[WIP] Remove unneeded JNI registrations.

Rather than registering all jni bindings at startup, only get references
to the class object for those files which require bindings. All others
are satisfied by exporting symbols which can be found automatically by
dalvik.

Based on systrace, looks to be 10-30ms improvement but it's far less
than std dev on cold start tests I've done.

This is incomplete as it currently requires exporting all symbols.

BUG=

[Yaron, 2013-06-28 00:22:12]

Not meant for real review but Maria was curious. As the CL description mentions,
it appears to be a small win. I can't get an accurate assessment of size
increase because of the gyp change which we'd have to do more work to avoid.

[bulach, 2013-06-28 00:45:53]

this is very neat, I think we should go for it..
adding torne for his thoughts.

there's one subtlety though, I think we can live with it but want to highlight:
right now, it's just impossible to call any native prior to "MainWithNative" (or
the equivalent in the different shells).. after this change, it'll be possible
to call into native whilst it's being loaded / not fully initialized..
again, not a big deal, but may lead to hard to track bugs should anybody start
taking "call into native" for granted at all points..

https://codereview.chromium.org/18161004/diff/1/base/android/jni_generator/jn...
File base/android/jni_generator/jni_generator.py (right):

https://codereview.chromium.org/18161004/diff/1/base/android/jni_generator/jn...
base/android/jni_generator/jni_generator.py:742:
__attribute__((alias("${NAME}")));
I think you can combine the attributes:
__attribute__((alias("${NAME}"), visibility("default")))

[Torne, 2013-06-28 00:48:54]

This looks pretty neat as long as we can sort the export issue and the size
impact is okay. Even if it's not a huge win it should always be a win,
especially if we add more JNI bindings :)

[bulach, 2013-06-28 00:53:24]

btw, in terms of size: we'll be able to remove the "kMethods" arrays, which
contains all method names... granted, it's a bit larger with the full package
name, but it shouldn't be too bad, I hope...