Main parquet text loader

[artemru]

What does this PR do? Please describe:

• Enable text dataloading from parquet dataset (shared at row groups levels)
• Adding small other fixes:

• adding split to text and jsonl dataloader reader to unify the interface
• keeping example shuffling for jsonl and parquet dataloader
• adding possibility to pass a different "text" column
• fixing npc at other dataloaders

Does your PR introduce any breaking changes? If yes, please list them:
No BC, only new feature

Check list:

• Was the content of this PR discussed and approved via a GitHub issue? (no need for typos or documentation improvements)
• Did you read the contributor guideline?
• Did you make sure that your PR does only one thing instead of bundling different changes together?
• Did you make sure to update the documentation with your changes? (if necessary)
• Did you write any new necessary tests?
• Did you verify new and existing tests pass locally with your changes?
• Did you update the CHANGELOG? (no need for typos, documentation, or minor internal changes)

Minimal working example

from fairseq2 import get_runtime_context
context = get_runtime_context()

from fairseq2.datasets.text_parquet import *
from fairseq2.recipes.common import load_text_tokenizer
from fairseq2.recipes.config import TextTokenizerSection

from fairseq2.datasets._config import LengthBatching
from fairseq2.gang import FakeGang
from fairseq2.device import CPU


seed = 42
max_num_tokens = 8192 * 2
max_seq_len = 8192
read_options = TextReadOptions(
        batching=LengthBatching(max_num_tokens),
        num_accumulate=1,
        num_prefetch=100,
        seed=seed,
        max_num_batches=100,
        extras={},
    )
tokenizer = load_text_tokenizer(context, TextTokenizerSection(name="llama3_2_3b"))
text_encoder = tokenizer.create_encoder(mode="default")

pdataset = ParquetTextDataset.from_path(path_to_your_parquet_dataset, name="text_parquet")
reader = pdataset.create_reader(split=None,
                                text_encoder=text_encoder,
                                      gang=FakeGang(CPU),
                                      pad_idx=None,
                                      options=read_options,
                                      min_seq_len=1,
                                      max_seq_len=max_seq_len)
from tqdm.auto import tqdm
bb  = []
for batches in tqdm(iter(reader)):
    bb.extend(batches)

[artemru]

@cbalioglu : pad_value is not documented, I hope I used it correctly

[artemru]

this will keep the previous behaviour by default

[zyaoj]

LGTM! Thanks for addressing the comments.