read_point_obs() slow for file_format = "obsparquet"

[andrew-MET]

read_point_obs() with file_format = "obsparquet" is much slower than simply reading the data with arrow::open_dataset()and dplyr::collect().

Data that show the problem are under my user on atos at $PERM/parquet

[andrew-MET]

Experimentation suggests this may be due to the writing creating too many parquet files - it may be better to partition less often to get bigger files. Slow file systems will struggle with too many small files.