Thread safety: per-task state rather than per-thread

[stevengj]

Closes #153.

Only a relatively small slowdown (5–30% for Dec64) for basic arithmetic, so a task-local rounding mode doesn't seem to be a huge concern. OLD CODE:

julia> @btime sum($(rand(Dec64, 1000)));
  45.614 μs (0 allocations: 0 bytes)

julia> @btime sum(n -> d64"0.9999"^n, 1:1000);
  933.454 μs (0 allocations: 0 bytes)

NEW CODE:

julia> @btime sum($(rand(Dec64, 1000)));
  60.292 μs (0 allocations: 0 bytes)

julia> @btime sum(n -> d64"0.9999"^n, 1:1000);
  961.635 μs (0 allocations: 0 bytes)

(The first benchmark should be the worst case, because it is only calling +, which is relatively cheap.)

[stevengj]

For some reason eps(d64"1.0") is allocating now — apparently the escape-analysis on the Ref isn't succeeding?

Other math functions like sin(d64"1.0") that also check the exception flags don't seem to allocate…

Update: fixed, just needed a bit of inlining.