Optimize the netcdf I/O for high-resolution DA

[gmao-cda]

Category: enhancement

Description:
Time profiling for 0.25 deg MOM6 DAS for ~17000 obs with 1000 cores on Discover:

### TIMER(INITIALIZE):      8.05      8.05
### TIMER(READ_OBS):      8.25      0.19
### TIMER(READ_ENS_MPI):     42.41     34.16
### TIMER(WRITE_GUES_MEAN_SPRD):     60.83     18.42
### TIMER(DAS_LETKF):     63.13      2.30
### TIMER(WRITE_ENS_MPI):    139.07     75.94
### TIMER(WRITE_ANALYSIS_MEAN_SPRD):    153.00     13.93
### TIMER(FINAL):    153.81      0.81

Read/write ensembles takes 110 seconds.
write bkgd/anal mean/sprd takes 32 seconds.
DA takes 2.3 seconds

Now each member is written by one process. Seek possibility of writing one member from several processes using PnetCDF