auto merge of #16626 : ruud-v-a/rust/duration-reform, r=brson

This changes the internal representation of `Duration` from

    days: i32,
    secs: i32,
    nanos: u32

to

    secs: i64,
    nanos: i32

This resolves #16466. Note that `nanos` is an `i32` and not `u32` as suggested, because `i32` is easier to deal with, and it is not exposed anyway. Some methods now take `i64` instead of `i32` due to the increased range. Some methods, like `num_milliseconds`, now return an `Option<i64>` instead of `i64`, because the range of `Duration` is now larger than e.g. 2^63 milliseconds.

A few remarks:
- Negating `MIN` is impossible. I chose to return `MAX` as `-MIN`, but it is one nanosecond less than the actual negation. Is this the desired behaviour?
- In `std::io::timer`, some functions accept a `Duration`, which is internally converted into a number of milliseconds. However, the range of `Duration` is now larger than 2^64 milliseconds. There is already a FIXME in the file that this should be addressed (without a ticket number though). I chose to silently use 0 ms if the duration is too long. Is that right, as long as the backend still uses milliseconds?
- Negative durations are not formatted correctly, but they were not formatted correctly before either.

src/libstd/io/net/unix.rs

@@ -61,7 +61,7 @@ impl UnixStream {
     /// Connect to a pipe named by `path`, timing out if the specified number of
     /// milliseconds.
     ///
-    /// This function is similar to `connect`, except that if `timeout_ms`
+    /// This function is similar to `connect`, except that if `timeout`
     /// elapses the function will return an error of kind `TimedOut`.
     ///
     /// If a `timeout` with zero or negative duration is specified then