Implement zero-copy parsing

[kmcallister]

This is based on #60 but with substantial changes. The biggest difference is that we only use shared buffers for the character runs found by pop_except_from. The majority of the remaining spans are single ASCII characters, which have their own fast path. Everything else is a String as before.

This branch also drops many of the micro-optimizations from #60. Unlike that PR, we leave the parser rules alone for the most part.

r? @Manishearth or @SimonSapin (general review)

r? @cgaebel (iobuf usage in tendril.rs)

Depending on the specific content and the I/O chunk size, this branch speeds up tokenization by up to a few percent. I did not see any significant performance regressions with sensible chunk sizes.

I have plans for further optimizations, including following up on the rustc bugs @cgaebel identified in #60.

The branch already achieves a significant drop in allocations and memory consumption:

(preliminary numbers)

Webapp spec, single page:

pre-zerocopy

846,520 allocs, 37,567,992 bytes allocated
maximum resident: 11,248 kB

zerocopy

95,690 allocs, 25,698,480 bytes allocated
maximum resident: 3,648 kB

Wikipedia (GotG from servo-static-suite)

pre-zerocopy

62,705 allocs, 2,299,424 bytes allocated
maximum resident: 3,948 kB

zerocopy

11,549 allocs, 2,355,984 bytes allocated
maximum resident: 3,596 kB

[cgaebel]

Isn't pos redundant with buf now, since Iobuf has its own cursor?

[kmcallister]

But the owned variant of a Tendril doesn't.

Eventually I would like to replace BufferQueue with a rope, as discussed elsewhere.

[cgaebel]

Good point!

[cgaebel]

Doesn't this mean that any long string of text (i.e. that spans multiple chunks) would have to fall back to a String? How big are chunks in Servo?

I guess that's a small price to pay for more efficient iteration.

[cgaebel]

Why not just use len32?

[cgaebel]

Iobuf usage looks good to me (modulo comments)! I like this remix.

[kmcallister]

Now #141.