Remove pandas dependency from meldataset.py

[netlinux-ai]

Problem

meldataset.py imports pandas and uses it for one operation: filtering data_list by speaker_id when sampling a reference clip. Pandas is otherwise unused.

This adds a heavyweight dependency for a single filter call. It also pulls in pandas's transitive pyarrow dependency, which is the source of compatibility friction:

• pyarrow's published wheels assume a v2-baseline x86 CPU (SSE4.1 in static initialisers); older CPUs cannot load them.
• pyarrow tracks newer Python release schedules aggressively, sometimes lagging behind by a release.
• pip install weight goes up by ~80 MB for one filter line.

Fix

Replace the single pandas usage with a list comprehension + random.choice:

# before
ref_data = (self.df[self.df[2] == str(speaker_id)]).sample(n=1).iloc[0].tolist()

# after
matching = [r for r in self.data_list if r[2] == str(speaker_id)]
ref_data = random.choice(matching)

Drops both pandas and pyarrow from the dependency graph. Unused self.df member is removed; unused import pandas as pd is removed.

Tested with

• Full fine-tune training run on PyTorch 2.7.0
• Reference-clip sampling behaviour confirmed identical: same uniform-random selection within speaker
• ~80 MB dep removal verified