fix(NODE-1935): Rework read state caching

[blind-oracle]

• Narrow down the cacheable paths to the following form:

  [
    [ "canister_ranges", "aaaaa-aa" ],
    [ "subnet", "aaaaa-aa" ],
  ]
  

  That is, only requests which have 2 paths, each of them has 2 labels, first of which is canister_ranges or subnet (or the other way around) and the second one is a valid Principal (in essence this just means shorter than 30 bytes). All other paths are deemed non-cacheable. This also fixes an issue - before we considered paths with at least one of the canister_ranges or subnet cacheable, where in fact we aim to cache only if we have both.

• Add an explicit check that the request is a subnet read state

• Move the logic that decides if paths are cacheable or not to process.rs to avoid storing the whole paths Vec in the request context, which might be large, up to 4MB. This avoids keeping the large context around for the duration of the request - only cacheable paths are stored. It's not ideal from architecture viewpoint, but I don't see other way w/o reworking the code a lot...

• Add early checks for the request body in preprocess and response body in read state caching middleware to avoid dealing with known-big in advance bodies

• Fix the weigher to size the paths correctly (not that it matters much anymore since we only cache small ones).

• Add/fix some tests

[Bownairo]

nit: Since this is pub and could be constructed directly, can we have a note that the paths should be sorted?