Fix language labels and anonymise source identifiers
Derived from the existing release; audio bytes and all annotations are unchanged.
language — use the full language name on every row. The 210 in-the-wild samples previously used ISO codes (bn, gu, hi, kn, ml, mr, or, pa, ta, te) while near- and far-field rows already used full names.
file_name → recording_id — remove file_name, which held the source filename (YouTube video id for in-the-wild). Add recording_id, an anonymised grouping key (<lang>_<nf|ff|itw>_<nnn>). 1164 clips come from 590 source recordings; 47 recordings contribute more than one clip.
audio.path — rewrite to <sample_id>.wav. It duplicated file_name on all 1164 rows, so dropping the column alone would have left source ids in the audio struct.
Private sample_id → original file_name mapping is kept offline and is not part of this PR.
refs/pr/2 ref