A handful of test infrastructure fixes needed once the queue runner and
builder use distinct physical Nix stores sharing the same logical
store dir.

- `HydraTestContext`: create the builder and central roots up front
  with `make_path` (the `local?root=...` store won't initialise
  otherwise), and give the builder its own `nix_state_dir` /
  `nix_log_dir` fields alongside the existing `nix_store_dir`.

- `QueueRunnerBuildOne`:
  - Set `NIX_STORE_DIR` when starting `hydra-builder` with a `TODO`
    note: the builder currently reads this to advertise its logical
    store dir to the queue runner; it should take the value from the
    store URI instead.
  - Pump the queue-runner `IPC::Run` harness alongside the builder
    and flush accumulated stderr on every poll, so when a test hangs
    we actually see the queue runner's logs instead of a blank
    "Builder stderr: ..." wall.

- `queue-runner/notifications.t`: the "prebuild into a binary cache"
  preamble uses `nix-build` / `nix copy` / `nix log` / `nix-store`
  directly, which inherits the central store URI. That store has
  physical != logical, so building derivations whose `$PATH` points
  at the host `/nix/store/...-coreutils/bin` fails with `mkdir: not
  found`. Point all of these at the builder store (which has physical
  = logical) via `--store`, and clean up the builder's log dir at
  the end of the subtest instead of the central one.

  This also makes sure that the test doesn't rely on `nix-build` etc.
  doing things to the central / queue runner's store that we didn't
  realize, so we don't accidentally rely on those things (e.g. the test
  would fail without them.)