Revert default in-memory for small datasets

[albertvillanova]

Users are reporting issues and confusion about setting default in-memory to True for small datasets.

We see 2 clear use cases of Datasets:

• the "canonical" way, where you can work with very large datasets, as they are memory-mapped and cached (after every transformation)
• some edge cases (speed benchmarks, interactive/exploratory analysis,...), where default in-memory can explicitly be enabled, and no caching will be done

After discussing with @lhoestq we have agreed to:

• revert this feature (implemented in Set default in-memory value depending on the dataset size #2182)
• explain in the docs how to optimize speed/performance by setting default in-memory

cc: @stas00 #2409 (comment)

[albertvillanova]

cc: @krandiash (pinged in reverted PR).