Be more defensive about inplace decompression in V2

[martindurant]

Fixes #889

[martindurant]

cc @miohtama

[miohtama]

I can confirm the fix works.

Unrelated, as I was working on it, I also timed both PyArrow and FastParquet read methods:

from pyarrow import parquet as pq  
from fastparquet import ParquetFile  

path = "./lending-reserves-all.parquet"

start = time.time()
df1 = pq.read_table(path).to_pandas()
print("PyArrow is", time.time() - start)

start = time.time()
pf2 = ParquetFile(path)
df2 = pf2.to_pandas()
print("FastParquet is", time.time() - start)

PyArrow is 6.349822044372559
FastParquet is 12.303039073944092

[martindurant]

For decimals, fastparquet currently converts to float and this probably accounts for the time difference. After conversion, float operations will usually be faster. my attempt at implementing decimals natively was even faster than anything that pyarrow can do, but due to lack of interest, I didn't implement direct fastparquet->decimal reading.

[miohtama]

@martindurant Thank you for the feedback.

Note that the difference is not that huge and likely depends on the data.