Add a cost keyword to @testitem and schedule by it

[jey]

Close #241

Adds a cost keyword to @testitem declaring roughly how long the item takes to run, in nominal seconds, and schedules by it: items that declare a cost are claimed before those that don't, most expensive first, so a long item can't start near the end of the run and leave the other workers idle waiting for it.

@testitem "slow integration test" cost=450 begin
    @test long_running_thing()
end

Details:

• The queue is sorted by (status_when_last_seen, -cost, number); number is unique, so the order is deterministic regardless of sort stability. This also makes the existing failures_first ordering deterministic, which it previously was not. failures_first takes precedence over cost within the key.
• Items with no declared cost sort after all items that do, in the order they would otherwise have run in. A run with no costs and no cached failure status takes the exact same path as before (unsorted queue, evenly spaced worker starts).
• When the queue is sorted, workers start at the front rather than at evenly spaced positions, so the most expensive items start immediately.
• A cost can also be a function of the run configuration, for items whose duration depends on it, e.g. cost=(cfg -> 15 + 450 / cfg.nworker_threads). It receives (; nworkers::Int, nworker_threads::Int) and is called exactly once per item, in the coordinator, before any item runs; the returned number replaces the function, so no user function is ever serialized to a worker. nworker_threads is the thread count items will actually run with (the setting, or Threads.nthreads() when serial).
• Costs are validated to be finite and non-negative (Bool rejected); a function must return a Real or nothing; errors while evaluating a cost name the test item.

Tests cover the macro parsing/validation, the sort (including ties, precedence with failures_first, and worker starting positions), cost-function resolution, and end-to-end run order via runtests with nworkers 0/1/2. README and docstring updated; version bumped to 1.36.0.

Questions:

• Happy to rename the keyword if cost isn't the right word.
• Should there be a run-level switch on runtests to ignore declared costs (keyword + env var, like failures_first)? As is, declaring one cost opts the whole suite into the sorted-queue regime on every run.

[jey]

Marking this ready for review since the two CI test failures appear to be spurious.