[bugfix][hail][oops] fix region reference counting in new aggregators

[catoverdrive]

I pulled out a flag to cache in java, but accidentally got rid of the thing it was actually doing.

This should be fixed now; with a smaller test mt I'm seeing the number of allocated regions be consistent between combOps:

...
2019-08-06 17:21:17 Hail: INFO: Region count for combOp
    regions: 27
     blocks: 28
       free: 25
       used: 2
2019-08-06 17:21:17 Hail: INFO: Region count for combOp
    regions: 27
     blocks: 28
       free: 25
       used: 2

[catoverdrive]

@chrisvittal I hit a segfault on this PR in the decoder trying to do

mt = hl.read_matrix_table(filename)
summarized = hl.experimental.densify(mt)

summarized.write(output1, overwrite=True)
summarized2 = hl.read_matrix_table(output1)

def test1():
	summarized._force_count_rows()

def test2():
	summarized2._force_count_rows()
	
print(timeit.repeat(test1, number=1, repeat=5))
print(timeit.repeat(test2, number=1, repeat=5))

The weird thing is that it appears to be sporadic, since it successfully ran all 5 iterations of test 1 and was on the last iteration of test2, judging by the number of spark stages, and it ran successfully when I just ran it a second time.

[chrisvittal]

Unused parameter.

addressed

[catoverdrive]

(re: segfault---I haven't hit this again, and the place where I hit it leads me to believe I might have just been doing something totally dumb---like rebuilding the jar in the middle of this)