Proper checks on nrows and chunksize for read_csv

[toobaz]

From #15755 (e.g. this comment).

Current status:

• there is a check on nrows, which ensure floats are converted to int with no loss of precision, but not that the resulting/passed value is non-negative
• there is no check on chunksize, so that a pd.concat based on chunksize=0 will enter an infinite loop

I guess we want to ensure chunksize > 0 and nrows >= 0.