When processing a node, miniaudio will read into a temporary buffer
before mixing input attachments. This commit removes the per-node heap
allocation and replaces it with a per-graph stack. This should result
in less memory usage at larger scales, but at the expense of slightly
more usage at smaller scales.

The size of the stack can be configured via ma_node_graph_config. If
ma_engine is being used, it can be done via ma_engine_config.