Fixes to eventual consistency of shared intervals, including significant LocalReference changes

[ransomr]

Implement new MergeTree LocalReference APIs to fix eventual consistency of SlideOnRemove LocalReferences.
Update SharedIntervals to use new APIs.
For more detailed documentation, see packages/dds/merge-tree/REFERENCEPOSITIONS.md

[Abe27342]

Looked at tests, not too many comments there. Obviously there are still some TODOs, but I'm ok with addressing them in follow-ups.

[Abe27342]

What's the rationale for the explicit update of reference offsets on segment removal? I don't think we can come up with an eventually consistent scheme while doing that. If it's just to make sure the observed offset matches the segment's position in the SharedString, there is already logic which handles returning 0 in getOffset() when the segment is removed.

Here's an explicit fuzz test case (minimized, so pretty readable) which demonstrates why this strategy loses information:

[
    {
        "type": "addText",
        "index": 0,
        "content": "ABCDEFGHIJ",
        "stringId": "B"
    },
    {
        "type": "synchronize"
    },
    {
        "type": "addInterval",
        "start": 6,
        "end": 8,
        "collectionName": "comments",
        "stringId": "B",
        "id": "7a5d683d-4b3e-44b3-8f91-e8f9d664de5a"
    },
    {
        "type": "addText",
        "index": 7,
        "content": "XYZ",
        "stringId": "B"
    },
    {
        "type": "removeRange",
        "start": 2,
        "end": 10,
        "stringId": "A"
    },
    {
        "type": "synchronize"
    }
]

A needs to be able to slide the references added by B to the start and end of XYZ, respectively, but it won't know how to do this if it blindly sets both offsets to 0 upon acking them.

(side note: I do think we will probably need to either expose a method which gets the offset ignoring segment removals, or push reference comparison + "split" knowledge somehow to the reference itself. There are a couple maintenance things which we do need to make sure we preserve that information)

[msfluid-bot]

⯅ @fluid-example/bundle-size-tests: +6.74 KB

Metric Name	Baseline Size	Compare Size	Size Diff
aqueduct.js	392.54 KB	395.14 KB	⯅ +2.6 KB
connectionState.js	711 Bytes	711 Bytes	■ No change
containerRuntime.js	211.42 KB	211.42 KB	■ No change
loader.js	146.35 KB	146.35 KB	■ No change
map.js	38.3 KB	38.3 KB	■ No change
matrix.js	121.7 KB	123.15 KB	⯅ +1.45 KB
odspDriver.js	144.89 KB	144.89 KB	■ No change
odspPrefetchSnapshot.js	36.83 KB	36.83 KB	■ No change
sharedString.js	138.32 KB	141.01 KB	⯅ +2.69 KB
Total Size	1.23 MB	1.24 MB	⯅ +6.74 KB

---

Baseline commit: f01e36f

Generated by 🚫 dangerJS against a7651f8

[Abe27342]

REFERENCEPOSITIONS.md could use another once-over, but generally this looks great. Thanks for all your hard work on it :)