Implements whatwg/html#6962 . Improves performance
when <meta charset> occurs in head but after the first kilobyte and aligns
behavior better with WebKit and Blink.

The main change is to avoid reloads when meta appears within head but
after the first kilobyte. Prior to this change, Gecko reloaded in that
case (in compliance with the spec!) even though WebKit and Blink did not.

Differences from WebKit and Blink:

* WebKit and Blink honor <meta charset> in <noscript>. This implementation
  does not.
* WebKit and Blink look for meta as if the tree builder was unaware of
  foreign content. This implementation is foreign content-aware. This
  makes a difference for CDATA sections that contain a > before the meta
  as well as style and script elements within foreign content. This could
  happen if the CDATA section that has mysteriously been introduced around
  a what looks like a meta tag also contains another prior tag-looking
  run of text.
* This implementation processes rel=preload and speculative loads that are
  seen before <meta charset> has been seen. WebKit and Blink instead first
  look for the meta and rewind before starting speculative parsing.
* Unlike WebKit, if there is neither an honored meta nor syntax resembling
  an XML declaration, detection from content takes place (as in Blink).
* Unlike Blink, if there is neither an honored meta nor syntax resembling
  an XML declaration, the detection from content is not dependent of network
  buffer boundaries.
* Unlike Blink, detection from content can trigger a reload at the end of
  the stream if the guess made at that point differs from the first guess.
  (See below for the definition of the input to the first guess.)

Differences from the old spec and Gecko previously:

* Meta inside script and RCDATA elements is no longer honored.
* Late meta is now ignored and no longer triggers a reload.
* Later meta counts as early enough meta: In addition to the previous
  meta within the first 1024 bytes, now a meta that started within the first
  1024 bytes counts as early enough. Additionally, if by then there hasn't
  been a template start tag and head hasn't ended, meta occurring before the
  earlier of the end of the head or a template start tag counts as early
  enough.
* Meta now counts as not-late even if the encoding label has numeric
  character reference escapes.
* Syntax resembling an XML declaration longer than a kilobyte is honored if
  there is no honored meta.
* If there is neither an honored meta nor syntax resembling an XML declaration,
  the initial chardetng scan is potentially longer than before: the first 1024
  bytes, the token spanning the 1024-byte boundary if there is such a token,
  and, if by then head hasn't ended and there hasn't been a template start tag
  until the end of the template start tag or the end of the token that causes
  head to end, ever comes first. However, if the token implying the end of the
  head is a text token, bytes only to the end of the previous non-text token is
  considered. (This definition avoids depending on network buffer boundaries.)
* XML View Source now uses the code for syntax resembling an XML declaration
  instead of expat for extracting the internal encoding label.

Reftest are added as both WPT and Gecko reftests in order to test both http:
and file: URL scenarios. The Gecko tests retain the WPT <link> tags in order
to use the exact same bytes.

An encoding declaration has been added to a number of old tests that didn't
intend to test the new speculation behavior especially in the context of
https://bugzilla.mozilla.org/show_bug.cgi?id=1727750 .

Differential Revision: https://phabricator.services.mozilla.com/D125808

bugzilla-url: https://bugzilla.mozilla.org/show_bug.cgi?id=1701828
gecko-commit: 9a8abd87cc7935f29b94248c1a6f8203faa14403
gecko-reviewers: smaug