Don't wait until all tasks have failed before outputting error text

[jyn514]

Due to Ekam's build model, it doesn't know whether a task will succeed later when more dependencies have finished building. As a result, it waits until all tasks have failed before outputting any errors. This is frustrating when the tasks take a long time to run (e.g. when one test out of several dozen fails), especially if the tasks are in fact independent.

I don't know if this can be solved in the general case, but #24 might help with the case where the tasks are independent since Ekam will know it doesn't need to wait for the other tests to fail.

[kentonv]

This is tough, because it's hard for Ekam to know if rules that run later might produce a new artifact which changes the result of a previous rule.

• We could maybe output the error if all dependencies were successfully resolved... but even that can theoretically be premature, as it's possible that an artifact generated later will be a more-exact resolution. For example, say someone has Cap'n Proto installed on their system, but is trying to build a new version from source... an include for <capnp/compan/byte-stream.capnp.h> might resolve to the system-installed copy early on, but the in-tree copy may be generated later, requiring the dependent to be rebuilt.

• We could perhaps use knowledge of the dependency tree from a previous build to decide when we think all dependencies are done, but this could lead to incorrect results whenever the dependency tree has changed between builds (e.g. the developer added a new include).