Revamp the core approach to converting Confluence to BCD

[foolip]

There were a number of issues here:

• Releases from different OSes were mixed and could lead to very strange suggested edits, especially for Safari.
• The version numbers suggested didn't always exist in BCD, sometimes because they weren't correctly trimmed.
• Patches which had both version_added and version_removed could be partially applied, which was the default because version_removed was only added with --remove.
• Safari for iOS versions were incorrect per BCD guidelines.

This was addressed by:

• Only using one browser+OS combination for each BCD browser.
• Getting the version number from BCD by matching components.
• patchPredicte considers the whole patch, not its parts.

Also, confluenceRowToCompatJsonFragment was reduced to a single scan through the row, which necessitated changing how the support object is updated. Important changes:

• Existing data is not used, that's left to patchPredicate.
• version_added is never set to true, but uses ≤ if necessary.
• Rows where support is true for every release of a browser are no longer skipped. (Was in the p => !!p.f(row) filter.)

The many repeated calls to getMdnBrowserRelease could be hoisted out to be less wasteful, but speed is not of the essence here.

Drive-by: replace use of foam.String.isInstance

[foolip]

I'm submitting this as a pull request for visibility. @vinyldarkscratch, this is in order to fix a bunch of bugs I found so that I could extract all the juice possible before https://github.com/mdn/browser-compat-data/pull/6435.

However, it ended up being an exercise in how to map a series of test results to BCD entries, and there may be something to learn/reuse here. In particular, a bit of an unsolved problem is how to deal with transient additions or removals, where they're reverted a release or two later. This results in a bunch of noise, but worse, if only the addition is trusted then the result will be wrong.

[foolip]

Self-reviewing and merging this. More bugs will likely be found while testing the sum of recent changes before release.