Status of CSET testing with Pangu and FastNet

[jwarner8]

Basing off branch fix_ml, testing all functionality with nearly all surface and pressure diagnostics on, using minimal data for Pangu and FastNet for surface and pressure level functionality (two short 48 forecasts with overlap).

- Pangu: all functionality, using SEQ.

1. Bake fail: domain_mean_y_wind_vertical_profile_seq.yaml, with traceback all the input array dimensions except for the concatenation axis must match exactly, but along dimension 1, the array at index 0 has size 28 and the array at index 1 has size 13. Initial suspicion of this problem is that UM has more pressure levels than Pangu. Happens for x and y wind (but not airT?)
2. Aggregate fail
   2.1. y_wind_pressure850_difference_um_-_pangu_aggregation_over_all_cases failed to merge into a single cube, where cube name differs cube.var_name differs: 'y_wind' != 'v_wind_on_pressure_levels' and cube.shape differs: (187, 142) != (281, 321). Seems OK for surface variables, so suspicion is there is some logic and extra steps missing in the pressure level recipes (including regridding for example onto common grid).
   2.2. histogram_air_temperature_aggregation_by_validity_time Failing with cannot use 'numpy.ndarray' as a set element - unsure but could be a lack of data.
   2.3. Same as bake fail, with vertical profile aggregation where all the input array dimensions except for the concatenation axis must match exactly, but along dimension 1, the array at index 0 has size 28 and the array at index 1 has size 13`

• Pangu: all functionality, using MEAN. (all bake worked, which could be useful for debugging why SEQ failed for bake where pressure levels differ).

1. Aggregate fail
   1.1. All failures cover the same cases as SEQ above (numpy, mismatch metadata, and pressure levels differing).

- FastNet: all functionality, using SEQ.

1. Bake fail: same challenges as Pangu above.
2. Aggregate fail: same challenges as Pangu above.

Combined run

All working, except for above, so solution is turning off spatial plev differences, and histogram aggregation on validity time.