Speed up PandasDataset further

[lostella]

Description of changes: More speedup following #2435. Testing on Python 3.8.13 with the data from #2363, using the following code:

from pathlib import Path
from time import time
import pandas as pd
from tqdm import tqdm

from gluonts.dataset.pandas import PandasDataset

df = pd.read_parquet(Path(__file__).resolve().parent / "long_df_sample.parquet")

t0 = time()
ds = PandasDataset.from_long_dataframe(
    dataframe=df,
    item_id="item_id",
    timestamp="timestamp",
    freq="M",
)
t1 = time()
print(f"construction time: {t1 - t0}")

N = 3

t0 = time()
for _ in range(N):
    for entry in tqdm(ds):
        pass
t1 = time()
print(f"average iteration time: {(t1 - t0)/N}")

Before the PR:

construction time: 0.3739509582519531
100%|█████████████████████████████████████████████| 25000/25000 [00:10<00:00, 2294.39it/s]
100%|█████████████████████████████████████████████| 25000/25000 [00:06<00:00, 3660.79it/s]
100%|█████████████████████████████████████████████| 25000/25000 [00:06<00:00, 3618.71it/s]
average iteration time: 8.413320620854696

After the PR:

construction time: 0.3727140426635742
100%|█████████████████████████████████████████████| 25000/25000 [00:08<00:00, 2825.26it/s]
100%|█████████████████████████████████████████████| 25000/25000 [00:05<00:00, 4383.24it/s]
100%|█████████████████████████████████████████████| 25000/25000 [00:05<00:00, 4380.71it/s]
average iteration time: 6.958552281061809

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

Please tag this pr with at least one of these labels to make our release process faster: BREAKING, new feature, bug fix, other change, dev setup

[lostella]

I think this requires improving tests before it can be merged: for each use case for PandasDataset, we should assert that each dataset element is exactly what we expect (field names, types, shapes).