Raise warning instead of error for block offloading with streams

[a-r-r-o-w]

Addresses: #11375 (comment)

Note that you could use num_blocks_per_group > 1 when using streams (the previous implementation allowed that, but only if the initialization order of blocks was the same as invocation order), but typically the computation of a single block completes before the next block finishes onloading, so setting it to 1 and only supporting that case seems reasonable.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.