Better tests for Float.parseFloat/Double.parseDouble, plus some fixes.

I'm not happy with the magic 1024s, and I had hoped to have one hack to
cover both harmony bugs. But this passes all our existing tests, all
harmony's existing tests, and the new tests I wrote while looking at
the code. (The new tests all pass on the RI too.)

I've also made a few doc comments slightly less unclear.

Change-Id: Ib72fd20bb3b157c7db5a09913247eed3cc746a2e
6 files changed