[WIP] Use a union to reduce the size of SmallVec

[Amanieu]

Uses a union to eliminate the space used by the enum determinant. The new layout looks like this:

struct SmallVec<A> {
    capacity: usize,
    union {
        inline: ManuallyDrop<A>,
        heap: struct {
            ptr: *mut A::Item, 
            len: usize,
        } 
    }
}

The capacity field is used to determine which of the two union variants is active:

• If capacity <= A::size() then the inline variant is used and capacity holds the current length of the vector (number of elements actually in use).
• If capacity > A::size() then the heap variant is used and capacity holds the size of the memory allocation.

Since unions with drop are still currently unstable, this is kept behind a "union" cargo feature, which falls back to an implementation using an enum if disabled.

---

This change is 

[Amanieu]

Hmm there seems to be a regression in the benchmark results. I will look into this.

 name                     old ns/iter  new ns/iter  diff ns/iter   diff %  speedup 
 bench_extend             188          328                   140   74.47%   x 0.57 
 bench_extend_from_slice  48           50                      2    4.17%   x 0.96 
 bench_from_slice         47           49                      2    4.26%   x 0.96 
 bench_insert             1,125        1,335                 210   18.67%   x 0.84 
 bench_insert_from_slice  94           104                    10   10.64%   x 0.90 
 bench_insert_many        402          385                   -17   -4.23%   x 1.04 
 bench_macro_from_elem    194          257                    63   32.47%   x 0.75 
 bench_macro_from_list    43           47                      4    9.30%   x 0.91 
 bench_push               382          351                   -31   -8.12%   x 1.09 
 bench_pushpop            282          671                   389  137.94%   x 0.42 

[emilio]

debug_unreachable may be worth in these.

[bors-servo]

☔ The latest upstream changes (presumably #93) made this pull request unmergeable. Please resolve the merge conflicts.

[eira-fransham]

You might want to recommit this without rustfmt. Ideally, the rustfmt would be a seperate PR.

Also, is the change from an == (fast) to a < (slower) worth the 1 byte saving (plus padding)? The benchmarks suggest that it is not.

[Amanieu]

I'm going to close this since I haven't had time to work on it lately.