feat: Create datasets with 24h accumulation fields

[lzampier]

Models with 24h timestep are right now trained on datasets with 6h frequence and thus accumulations of 6h. This means that every 24h we predict the rain that falls in the last 6h os the timestep. We want to takle this from an anemoi-datasets side, to build ready to use datasets with 6h accumulations. This requires modifying the accumulation classes because multiple mars requests are needed to accumulate over 24h the 06 and 18 ERA5 forecasts, each with 18h lead time.

Ultimately, for S2S purposes, we need a dataset similar to https://anemoi.ecmwf.int/datasets/aifs-ea-an-oper-0001-mars-o96-1979-2023-6h-v8?tab=recipe

With only the accumulated parameters.
With 24 accumulations (freq is still 6h)

We want to use the latest forecast to do the accumulation, i.e. with minimal value for step. For instance:

For 00Z, accumulate on 24h as follow:
0-6 : use step 6 to 12 of the forecast starting at 18 (two days before)
6-18 : use step 0 to 12 of the forecast starting at 6 (one day before)
18-24 : use step 0 to 6 of the forecast starting at 18 (one day before)

For these parameters:

• tp
• lsp
• cp
• sf
• lsf
• csf
• ro
• ssrd
• strd

@floriankrb @stietsche