bytes: iobuf_copy requests large contiguous ranges

[jcsp]

(Internal slack ref https://redpandadata.slack.com/archives/C01ND4SVB6Z/p1641835943108600)

• iobuf_copy rounds required bytes down to next pow2 to choose allocation size. Means if the copy size is big, so is the allocation.
• it's intentional that copy doesn't aggressively fragment the data, to avoid generating fragmentation that would hurt other subsystems that need contiguous ranges. However, this also means that copy is very sensitive to fragmentation, and will throw bad_alloc on systems with somewhat fragmented allocators.
• Large calls to iobuf_copy originate in the kafka fetch path, where each partitions read_result structure has a buffer whose size is rougly the max_size the client requested (10MB default). Consequently iobuf_copy tries to do 8MB contiguous allocations.

[jcsp]

First cut of backoff on bad_alloc is in #3427, for the si-testing branch where the issue is cropping up in testing.