Use groupby in for loop instead of filter in generate_balance_sample

[crispy-wonton]

Groupby is more efficient than filtering. So move groupby to weighting loop in script (instead of filtering in generate_balance_sample function and test on EPC sample subset.

Original comment:
I'm not sure if it's worth checking and changing (I don't think it is at this stage), but FYI that I found in the evaluation that a groupby loop was more efficient than a loop in which you subset the df per lsoa within the loop. i.e.

df_select = df.select(cols) # the groupby was also faster with less columns, so worth doing this first too
for lsoa, sample in tqdm(df_select.group_by("lsoa")):
      results = generate_balance_sample(sample, target_marginals, lsoa)

Originally posted by @lizgzil in #26 (comment)