Track the cost of a subgraph

[lutter]

It would be useful to measure how much work a subgraph actually causes, e.g., so that we can make decisions on how to balance indexing of subgraphs across different graph-node instances.

To start, we should, for each block, look at the time that a subgraph spends running handlers plus the time it takes to store entities in the database as the cost for each block, and then compute a longer-term average using exponential decay, i.e. if the cost for one block is b_i, make the subgraph cost sum_{i=0}^n c^(n-i) * b_i for some c < 1; that measure would be easy to compute on a running basis, and the choice of c would determine over how many blocks we average the cost.

[leoyvens]

An important thing to define is if we want to measure the cost of syncing, which is done in block ranges, or the cost of updating which is done block by block.

It'd be great if we could time and register not only the total cost for a block (or range) but also the cost of each processing stage (after we rationalize what these stages are), this would help a lot in informing the investigation in #856 on how to speed up syncing.