grib_tree function fails to properly handle ECMWF ensemble data

Issue Description

The grib_tree function in kerchunk.grib2 doesn't properly handle ECMWF ensemble forecast data, specifically:

1. It fails to recognize and preserve ensemble member information
2. It significantly reduces the number of groups in the output compared to input
3. It doesn't provide a way to access ensemble dimension in the resulting zarr structure

Reproduction

When processing ECMWF ensemble data with 19 variables and 51 ensemble members (969 total message groups):

from kerchunk.grib2 import scan_grib, grib_tree
import datatree
date_str='20240229'
ecmwf_s3url=f"s3://ecmwf-forecasts/{date_str}/00z/ifs/0p25/enfo/{date_str}000000-0h-enfo-ef.grib2"
esc_groups = scan_grib(ecmwf_s3url)
original_tree = grib_tree(esc_groups)
gfs_dt = datatree.open_datatree(
    fsspec.filesystem("reference", fo=original_tree).get_mapper(""), 
    engine="zarr", 
    consolidated=False
)

# The key test: can we access ensemble members?
print(gfs_dt.keys())  # Check for variables

The resulting structure loses ensemble information, making it impossible to distinguish between different ensemble members in the output.

This gist explains the situation and a wayforward to have the ensemble number in the grib_tree.