Iterable dataset map with explicit features causes slowdown for Sequence features

[alex-hh]

Describe the bug

When performing map, it's nice to be able to pass the new feature type, and indeed required by interleave and concatenate datasets.

However, this can cause a major slowdown for certain types of array features due to the features being re-encoded.

This is separate to the slowdown reported in #7206

Steps to reproduce the bug

from datasets import Dataset, Features, Array3D, Sequence, Value
import numpy as np
import time
features=Features(**{"array0": Sequence(feature=Value("float32"), length=-1), "array1": Sequence(feature=Value("float32"), length=-1)})
dataset = Dataset.from_dict({f"array{i}": [np.zeros((x,), dtype=np.float32) for x in [5000,10000]*25] for i in range(2)}, features=features)

ds = dataset.to_iterable_dataset()
ds = ds.with_format("numpy").map(lambda x: x)
t0 = time.time()
for ex in ds:
    pass
t1 = time.time()

~1.5 s on main

ds = dataset.to_iterable_dataset()
ds = ds.with_format("numpy").map(lambda x: x, features=features)
t0 = time.time()
for ex in ds:
    pass
t1 = time.time()

~ 3 s on main

Expected behavior

I'm not 100% sure whether passing new feature types to formatted outputs of map should be supported or not, but assuming it should, then there should be a cost-free way to specify the new feature type - knowing feature type is required by interleave_datasets and concatenate_datasets for example

Environment info

3.0.2