Fix off-by-one error in check_all's capacity check

[antifuchs]

This fixes #11:

I'd missed that the "weight" variable is additional weight that gets
added to the first cell, so check_all (when given one more element
than fit in the burst capacity) would return a "denied" result instead
of an "insufficient capacity" result - effectively preventing futures
from ever completing.

Now, we treat additional weight as what it is:

• rename "weight" to "additional_weight" to make it clear what is
  going on
• Add a comment over the capacity check, clarifying the off-by-one
  error potential
• Add a test about for more capacity-excession scenarios, including
  the off-by-one, to prevent regressions.

Thanks to @jean-airoldie for discovering the bug & reporting it!

[antifuchs]

Hrm, for some reason the bench task on circleci is failing to run to completion (or even start - it fails to compile criterion): I tried setting the timeout to 20min, but it's still exceeding that timeout. I'm guessing this is either circleci or the rust version that came in docker images is broken atm... we can retry tomorrow.

[antifuchs]

bors r+

(I had to bump the bench version to nightly, but that at least can compile the criterion crate. Reported as bheisler/criterion.rs#377).

[bors]

Build succeeded

• ci/circleci: ci_success