Rework DCP connection buffer sizing

Currently the DCP connection buffer size is configured to be (1% of JVM max heap for JVM max heap >= 8GiB, 10 MB otherwise) / <number of iodevices on the node>)

This can end up using a lot of the heap when there many link+buckets connected, and if there are many KV masters.

Instead of being % based for larger heaps, this will be configured as a static value, 16MiB, which will be distributed across all the KV connections in a compute partition. e.g. for a node w/ 4 compute partitions, each link+bucket will allow a maximum of 16*4 or 64MiB of connection buffer for DCP.

For heaps <8GiB, configure 10MiB/the number of compute partitions on the node instead of 16MiB.